Geostatistics for Mineral Resources

Uncertainty Assessment

A Short Course

October 2016

Clayton V. Deutsch

Centre for Computational Geostatistics

School of Mining and Petroleum Engineering
University of Alberta • Edmonton, Alberta, Canada
 Geostatistics for Mineral Resources
Uncertainty Assessment
This course will cover essential theory and practice of geostatistical modeling for resource and
reserves uncertainty; advanced material is covered at an introductory level. Important details
of simulation will be discussed and presented. Important current subjects of data of different
quality, combined categorical and continuous variable modeling, multivariate, geometallurgical
modeling and managing multiple realizations will be covered.

Day One 01 – Introduction and Overview

02 – Statistics and Declustering
03 – Variogram Calculation
04 – Variogram Modeling
05 – Theory of Kriging
06 – Practice of Kriging

Day Two 07 – Principles of Simulation

08 – Categorical Variable Modeling
09 – Multivariate Gaussian Distribution
10 – Gaussian Simulation
11 – Practice of Gaussian Simulation
12 – Uncertainty and Sensitivity Analysis

Day Three 13 – Overview of Multivariate

14 – Multivariate Simulation with Decorrelation
15 – Principles of Geometallurgical Modeling
16 – Model Post Processing and Data Spacing
17 – Classification
18 – Case Studies

This schedule will be modified to meet the objectives and pace of the instructor/participants

Copyright, 2016, Centre for Computational Geostatistics

All rights reserved. The entire set of course notes can be reproduced and distributed in their entirety
(with this notice intact). Small excerpts can be distributed with acknowledgement given to the source.

1
Geostatistics for Mineral Resources Uncertainty Assessment

Introduction and Overview

• Preliminary remarks and background

• Context of the course
• General notation and theory

Logistical Details

• Notes and reference material:

– Electronic version provided
– Notes can be printed if you want
– Supplementary reference material provided

• Requests:
– Ask questions right away
– Be on time
– No unauthorized recording

• Software
– Software is essential
– There are commercial alternatives
– This course is not on software

Lec 01 – page 1
 Background
• Geostatistics is an approach and a toolkit that applies statistical and numerical
analysis principles to geological variables
• The focus of this course is the fundamental principles of:
– Geological heterogeneity modeling
– Uncertainty assessment
– Decision-making
• Participants should leave with:
– Appreciation for the place of geostatistics in support of geological modeling
– Knowledge of the limitations of geostatistics
– Background necessary to dig into more details
• This course conveys selected concepts from geostatistics. Time is limited:
– Citation Program in Applied Geostatistics (four weeks plus project)
– Masters (M.Eng. and M.Sc.) Degrees (two years plus thesis)
– Doctorate (Ph.D.) Degree (two more years plus thesis)
• Background knowledge in statistics and mathematics would make this short
course easier, but background material will be given as we proceed
• Review the breadth of modern geostatistics drilling into some important
principles along the way
3

Numerical Modeling
• Historically, science involved (1) extensive data collection and physical
experimentation, then (2) deduction of laws consistent with the data
• Now, science is much more concerned with (1) understanding and
quantifying physical laws, and (2) numerical modeling for inference
• We now accept that uncertainty cannot be removed (account of E. Teller’s
statement of how science has changed)
“then we believed that everything could be predicted, now we
know that we can only predict in a probabilistic sense”

• In general:
– Numerical modeling has become more important than physical experimentation,
– Inductive reasoning has become more popular than deductive reasoning,
– Uncertainty is quantified and managed rather than ignored.
• Numerical modeling is ubiquitous in modern
science and engineering (virtually all design
is on the computer…)

• Given how important numerical modeling is,

we should study how to do it well
4

Lec 01 – page 2
 Historical Perspective
• A gambler's dispute in 1654 led to the creation of a
mathematical theory of probability by two French
mathematicians, Blaise Pascal and Pierre de Fermat

• Rev. Thomas Bayes (1702-1761) first used

probability inductively and established a
mathematical basis for probability inference
(ABCs of probability…)

• In 1812 Pierre de Laplace (1749-1827)

introduced a host of new ideas and
mathematical techniques in his book,
Théorie Analytique des Probabilités

• Carl Friedrich Gauss (1777-1855) developed the

important method of least squares and the
fundamental laws of probability distributions

History of Geostatistics
• D. Krige and H. Sichel studied reserve estimation problems in South
Africa from the 1950's establishing the problem
• Professor Georges Matheron (1930-2000) built the major concepts of the
theory for estimating resources he named Geostatistics
• The monumental Traité de géostatistique appliquée (Editions Technip,
France, 1962-63) defines the fundamental tools of linear geostatistics:
variography, variances of estimation and dispersion, and kriging.

• Two of Matheron's first students (Journel and David) would leave for the
USA and Canada and start new centers of geostatistical research
6

Lec 01 – page 3
 Geostatistics
• Business Need: make the best possible decisions in presence of
uncertainty. One of the biggest uncertainties is the numerical
description of the subsurface.

• Statistics is concerned with scientific methods for collecting, organizing,

summarizing, presenting and analyzing data, as well as drawing valid
conclusions and making reasonable decisions on the basis of such
analysis

• Geostatistics is a branch of applied statistics that places emphasis on:

(1) Geological context of the data
(2) Spatial relationship between the data
(3) Data of different volumetric support and precision

Some Preliminary Remarks

• Geostatistics is statistical modeling and not geological modeling
– Geological (physical) constraints must be considered
– Adapt techniques to mimic geological features

• Geostatistics is useful for:

– Putting geology into numbers
– Estimation
– Quantifying uncertainty
– Sample design
– Simulation / risk analysis

• Geostatistics does not:

– Work well as a black box
– Replace the need for common sense and good judgment
– Save time
– Make modeling any easier

Lec 01 – page 4
 Why Geostatistics?
• The best approach to model the spatial distribution of geological
properties at the present time
• Better modeling of variability:
– Controllable degree of spatial variability
– Estimates are more reliable
• Framework to integrate data:
– Geological interpretation
– Direct measurements (hard data) and secondary variables (soft data)
– Data representing different measurement supports
• Assess uncertainty in process performance due to uncertainty in
geological model
• Practicality / consistency with data
• Repeatability / audit-trail
• Easy to merge incremental data

Statistical Modeling of Geology

• There is a single true distribution of properties in each deposit at any instance
in geological time
• The true distribution is the result of a complex succession of physical, chemical,
and biological processes
• Although some of these depositional and diagenetic processes may be
understood, we do not completely understand all of the processes and have no
access to the initial and boundary conditions in sufficient detail to provide the
unique true distribution
• We adopt a numerical geostatistical model because we have no alternative

Rock Type 1
Rock Type 2
Rock Type 3
Rock Type 4
Rock Type 5
Rock Type 6

10

Lec 01 – page 5
 Stochastic Modeling
Reality Model
Distribution of Rock/Fluid Properties Distribution of the Rock/Fluid Properties

single true distribution multiple stochastic models

Recovery Process Recovery Process

actual process implemented numerical model of process

Field Response Field Response

Single true response Distribution of

possible responses
11

Uncertainty and Mapping

• Uncertainty: All numerical models would be found in error if we were to
excavate that unsampled volume and take exhaustive measurements: There is
uncertainty. This uncertainty exists because of our ignorance/lack of
knowledge. It is not an inherent feature of the deposit.

• Uniqueness and Smoothing: Conventional mapping algorithms were devised to

create smooth maps to reveal large scale geologic trends; for most
applications, however, the extreme high and low values often have a large
affect on process performance and reserve calculations

Transfer
Function

Distribution of possible
Multiple stochastic models responses

12

Lec 01 – page 6
 Some Comments on Uncertainty
• Uncertainty exists because of incomplete data:
• Cannot be avoided,
• Can be reduced by consideration of all relevant data, and
• Can be managed

• The main steps we will propose are:

1. Conceptualize a numerical modeling scheme
2. Quantify the uncertainty in each of the inputs
3. Construct numerical geological models
4. Transfer the input uncertainty through to output uncertainty
5. Make optimal decisions in presence of uncertainty

13

Review of Main Points

• Course will review the fundamentals of geostatistics
• Discuss the philosophy and specific tools
• Modern science and engineering involves:
– Inductive reasoning based on principles and limited data
– Extensive numerical modeling
• Uncertainty exists because of our incomplete knowledge
• Geostatistics does not make modeling easier, but value is added
with a good model of variability and uncertainty
• Model variability that affects resources/reserve
• Quantify uncertainty in numerical geological model
• Transfer geological uncertainty through to response uncertainty
• Must also quantify consequences of making a mistake for optimal
decision making

14

Lec 01 – page 7
Geostatistics for Mineral Resources Uncertainty Assessment

Statistics and Declustering

• Comments on coordinates and data

• Statistics
• Probability distributions
• Declustering

Stratigraphic Coordinates

• Facies and properties are Proportional

modeled in flattened stratigraphic

coordinates to account for:
– Differential compaction
– Structural deformation/faulting Truncation Erosion

– Erosion
– Filling preexisting topography
• Handled automatically in most
software
Onlap

Erosion
Combination

Lec 02 – page 1
 Vein Coordinates
• Define the hangingwall and footwall by surfaces
• Choose a gridding system within
• Assign grades
• Convert back to regular 3-D system

More Prerequisites
• A geological understanding of the site being modeled is essential

• A numerical framework must be setup to represent the relevant

geological structures

• Understanding the geology, data and numerical framework requires

flexible visualization and summarization of the numerical data
– Creative use of statistics can help, but this is a job for a human
– Careful review of all data and resulting model is essential

• Following are a number of probabilistic concepts used throughout

geostatistics and modern science/engineering

Lec 02 – page 2
Cumulative Distribution Function (CDF)

• The universal approach to represent uncertainty in a continuous

variable
• Replace unknown truth by a random variable Z
• Cumulative distribution function (CDF) of an RV is defined as:

F ( z )  Prob{Z  z}

• Properties of the CDF:

– F(z) non decreasing
– F(z)  [0,1]
– F(-) = 0 and F() = 1

• Three ways to get a CDF:

– Degree of belief
– Proportions from data
– Mathematical model 5

Probability Density Function (PDF)

• The probability of Z occurring in an interval from a to b (where b>a) is
the difference in the cdf values evaluated at points b and a:

Prob{Z  [a, b]}  F (b)  F (a)

• Probability density function (pdf) is the derivative of the cdf, if it exists:

F ( z  dz )  F ( z )
f ( z )  F ' ( z )  lim
dz  0 dz

• The cdf can be obtained by integrating the pdf (fundamental theorem

of calculus): z
F (z) 

 f ( z ) dz

• CDF is used most of the time, but it may be convenient to use PDF
6

Lec 02 – page 3
 Distribution Models
• Parametric distribution models have an analytical expression for the
probability (cdf or pdf), which is completely determined by a few
parameters
– The normal or Gaussian distribution
is the most important to us (see later):

– Parametric models sometimes relate to an underlying theory

• Normal distribution is the limit distribution for the central limit theorem
• Weibel distribution is used in reliability theory

• Nonparametric distributions come from the data themselves

– Directly from proportions less than thresholds
– There are often enough data
– Geological data do not follow a convenient parametric model

• It is possible to transform univariate distributions

7

Quantiles
• The p-quantile of the distribution F(zp) is the value zp for which:

F ( z p )  Prob{Z  z p }  p  [0,1]
Thus, the quantile can be expressed in an inverse form of the CDF
q(p) = F-1(p)  Inverse, not 1/--
• The probability value lends meaning to a particular number

first quartile

90% probability interval 8

Lec 02 – page 4
 The Gaussian Distribution
• The Gaussian distribution is a constant of the universe.
– AKA normal distribution, error function or the curve
– Unparalleled mathematical tractability – particularly in high dimensions
• Central Limit Theorem: the sum of a great number of independent equally
distributed (not necessarily Gaussian) standardized random variables tends
to be normally distributed

z  mz 1  y 2 /2
• The probability density: y f ( y)  e
z 2

Normal Score Transform

• Data are usually not Gaussian or normal, but we may want them to be
• Very common for continuous variables
• Steps:
1. Assemble F(z) to represent the Z-variable
2. Transform by matching quantiles, that is, y = G-1(F(z)) where G(y) is the standard
Gaussian CDF

• This standard approach is also called a Gaussian anamorphosis (that may

involve a second step to fit the normal scores transform with hermite
polynomials)
10

Lec 02 – page 5
 Expected Values
• The expected value is a statistical operator that is the probability weighted
average of the random variable: 
E{Z }   z  f ( z ) dz


E{Z 2 }  z  f ( z ) dz
2




• Expected value of a constant is the constant
 
E{a}   a  f {z} dz  a  f ( z ) dz  a
 

=1
• The expected value is a linear operator:
E{aZ }  a E{Z }
E{ X  Y }  E{ X }  E{Y }
11

The First Two Moments

• The expected value of a random variable is the mean or first moment:

m  E Z 
• The mean is the correct effective value for most variables
• The variance is a second order moment defined as:

Var{Z }   2  E   Z  m   E  Z   m
2 2 2

• The variance is a measure of the spread of the data from the mean.
• The standard deviation is the square root of the variance

• The dimensionless coefficient of variation (CV) is the ratio of the standard

deviation over the mean (/m)
• Conventional wisdom:
– CV<0.4 – not that variable – potentially straightforward
– CV>2.0 – quite variable – potentially difficult – consider subdividing

12

Lec 02 – page 6
 Representative Statistics
• Account of Dr. George Gallup

• The data are of unquestioned importance to us

• The challenge is to model the vast unsampled region between data

• This lecture is not about biased data or measurement errors

• This lecture relates to the representivity of the data

13

Spatial Location of Data

• Data are rarely collected to be statistically representative
• These “data collection” practices should not be changed; they
lead to the best economics and the greatest number of data in
portions of the study area that are the most important
• There is a need, however, to adjust the histograms and summary
statistics to be representative of the entire volume of interest

• Declustering techniques assign each datum a weight based on

closeness to surrounding data
– wi, i=1,…n
– histogram and cumulative histogram use wi, i=1,…n instead of 1/n
• Debiasing uses a secondary variable or trend to establish
representative histogram or proportions

14

Lec 02 – page 7
 Representative Statistics
• Historical mapping algorithms correct for preferential sampling: no
need for declustering in inverse distance, OK, …
• There is a need for representative proportions and histograms in
modern geostatistics:
– Checking models
– Global resource assessment
– As an input to simulation
• Simulation does not correct for preferential sampling even though
kriging is used inside simulation
• Cokriging with a secondary data does not correct the distribution:
correlation with the rank order of the data is used – the conditional
distributions are not used directly

15

Principle of Declustering
• Assign differential weights to data based on proximity to other data
• Note the example below for five data equal weighted (left) and
weighted (right)

w z
n

w z z
2
i i
i i
z i 1
n s i 1

w
n

i 1
i
w i
i 1 16

Lec 02 – page 8
 Example of Declustering
• Location map of 122 wells. The gray scale shows the underlying “true”
distribution of porosity (inaccessible in practice)
• Histogram of 122 well data with the true reference histogram shown as the
black line (inaccessible in practice)
• Note the greater proportion of data between 25 to 30 % and the sparsity of data
in the 0 to 20% porosity range

Number of Data 122

mean 22.26
0.12
40 std. dev. 5.58
coef. of var 0.25
maximum 29.31
35
upper quartile 26.28
median 24.34
30 lower quartile 20.03
0.08 minimum 6.06

Frequency
25

20

true distribution
15 0.04

10

0.00
0
0.0 10.0 20.0 30.0 40.0

Porosity, %
17

Declustering by Area of Influence

• Location map of 122 wells with polygonal areas of influence
• Declustering weights are taken proportional to the areas
• Weight between 25 to 30 % has been decreased and the weight in the 0 to 20%
porosity range has been increased
• Weights are very sensitive to the border. The true correct border is used in this
example.
Number of Data 122
mean 19.44
std. dev. 6.90
0.08 coef. of var 0.35
maximum 29.31
upper quartile 25.31
median 21.03
0.06 lower quartile 14.64
minimum 6.06
Frequency

polygonal weighting

0.04

0.02

0.00
0.0 10.0 20.0 30.0 40.0

Porosity, %

18

Lec 02 – page 9
 Cell Declustering
• Polygonal declustering works well when the borders / limits of the area
of interest are well defined
• Another technique, called Cell Declustering, is more robust in 3-D and
when the limits are poorly defined:
– divide the volume of interest into a grid of cells l=1,…,L
– count the occupied cells Lo and the number in each cell nl, l=1,…, Lo
– weight inversely by number in cell (standardize by Lo)

weights = 1/(2 x 33) = 0.0152

weights = 1/(5 x 33) = 0.0061

weights = 1/(4 x 33) = 0.0076

weights = 1/(2 x 33) = 0.0152

19

Cell Declustering
• Fixing the cell size and changing the origin often leads to different declustering
weights. To avoid this artifact
– A number of different origin locations are considered for the same cell size
– Weights are averaged for each origin offset
• Cell size should be the spacing of the data in the sparsely sampled areas
• The declustered mean is often plotted versus the cell size for a range of cell
sizes – a diagnostic plot that may help get the cell size

23.0 Number of Data 122

mean 20.02
0.10
std. dev. 6.63
coef. of var 0.33
maximum 29.31
22.0 0.08 upper quartile 25.32
Declustered Mean, Porosity, %

median 22.02
lower quartile 16.22
minimum 6.06
Frequency

0.06 cell declustering

21.0

0.04

20.0
0.02

minimum declustered mean

0.00
19.0
0.0 10.0 20.0 30.0 40.0
0. 200. 400. 600. 800. 1000.
Porosity, %
Cell Size
20

Lec 02 – page 10
 Cell Declustering: Some Notes
• Perform an areal 2-D declustering when drilling is vertical
• Consider 3-D declustering when there are horizontal or highly deviated
wells that preferentially sample certain stratigraphic intervals
• The shape of the cells depends on the geometric configuration of the
data - adjust the shape of the cells to conform to major directions of
preferential sampling
• Choosing the cell size
– Choose the cell size so that there is approximately one datum per cell in the
sparsely sampled areas
– When there are many data and it is known that the high- or low-valued
areas have been oversampled then the cell size can be selected such that
the weights give the minimum (or maximum) declustered mean of the data
– Consider 10-20 origin offsets to stabilise the results relative to the origin

21

Declustering
Mean Standard Deviation
Reference 18.72 7.37
Equal Weighted 22.26 +18.9% 5.58 -24.3%
Polygonal Declustering 19.44 +3.8% 6.90 -6.4%
Cell Declustering 20.02 +6.9% 6.63 -10.0%

• Polygonal declustering led to statistics closer to the reference histogram

in this 122 well example
• It has been observed that polygonal declustering works well when the
limits (boundaries) of the volume of interest are well defined and the
polygons do not change in size by more than a factor of, say, 10 (largest
area / smallest area)
• An alternative consists of performing a global estimation and
accumulating the weight that each datum receives (clustered data will get
less weight)
• Declustering only works when there are enough data to assign greater
and lesser weights - what do we do when there are too few data?

22

Lec 02 – page 11
 Review of Main Points

• Stationarity
• Probability distributions
• Summary statistics
• Build weighted distribution / summary statistics:
– Polygons of influence
– Cell declustering
– Global estimation (accumulate weight to each data)
• Calibrate to secondary data when weighting is deemed inadequate:
• Other considerations / comments:
– No recipe for correct application
– Want to go beyond a limited sample to the population
– Must decluster facies proportions
– Future geostatistical analysis will depend heavily on simple statistics
inferred early in the modeling efforts

23

Lec 02 – page 12
Geostatistics for Mineral Resources Uncertainty Assessment

Variogram Calculation

• Spatial Dependence
• Variogram and Covariance
• Calculation of the Variogram
• Basic Interpretation Principles

Spatial Correlation
• The three maps are remarkably
similar: all three have the same 140
data, same histograms and same
range of correlation, and yet their
spatial variability/continuity is quite
different

• The spatial variability/continuity

depends on the detailed distribution
of the petrophysical attribute (,)

• The charts on the left are

“variograms”

Lec 03 – page 1
 The Variogram

2 ( h)  E{[ Z ( u)  Z ( u  h)]2 }

• The Variogram calculated for lag distance h, corresponds to the

expected value of squared difference:
• Calculate for a set of lag distances to obtain a continuous
function.

Plot as a Function of Distance

• Variogram as a function of
distance in different directions

Each point corresponds to an h-scatterplot

4

Lec 03 – page 2
 Variogram Lags

• Consider data values separated by lag vectors

Tail
Tail

Head Head

Covariance and the Variogram

• Under stationarity the variogram, covariance and correlation

coefficient are equivalent tools for characterizing two-point correlation:

 ( h)   2  C ( h )
  2 [1   ( h)]

C (h)  E{Z (u)  Z (u  h)}  E{Z (u)}  E{Z (u  h)} C (0)   2

• Stationarity entails that:

m(u)  m(u  h)  m  E{Z}, u  A
Var(u)  Var(u  h)   2  Var{Z}, u  A

Lec 03 – page 3
 Covariance / Variogram

< 0

C(0) = 2
(h)
= 0

> 0

Distance “lags”
0 h
C(h)
Covariance is more intuitive,
but less commonly used due to
historical reasons.

Must plot variance to interpret variogram:

No correlation at variance  covariance = 0
Variogram points above sill imply negative correlation 7

Calculating Variograms

• Calculate the experimental variogram for distances and directions

where we have data
• Regularly gridded data (secondary data or model checking) are
straightforward – offsets of the grid node spacing

• Application
– Secondary data
– Checking models

Lec 03 – page 4
 Calculating Variograms

• Irregularly or variably spaced data will require some tolerance

• Angle – clockwise from North Y axis (North)

• Angle tolerance
• Bandwidth

• Dip angle – negative down Azimuth

• Angle tolerance Bandwidth

• Bandwidth
Azimuth
tolerance

• Unit lag distance X axis (East)

• Lag tolerance

Trade off Precision and Stability

10

Lec 03 – page 5
 Choosing Directions
• Consider the vertical (or down hole) variogram first
• Consider the omnidirectional (all directions taken together) next
• Look at maps and sections to identify directions – consider the
orientation of geological site and/or the data
• Create a “neutral” model using IDS or kriging
• Consider multiple directions before choosing a set of 3
perpendicular directions (major horizontal direction & two
perpendicular to major direction)

11

Lag Distance and Tolerance

• Lag separation distance should coincide with data spacing
• Lag tolerance typically chosen to be ½ lag separation distance
– in cases of erratic variograms, may choose to overlap calculations so lag
tolerance > ½ lag separation.
• The variogram is only valid for a distance one half of the field size 
choose the number of lags accordingly
• Compute the vertical variograms in one run and the horizontal
variograms in another
Y axis (North)

Azimuth
Bandwidth

Azimuth
tolerance

X axis (East) 12

Lec 03 – page 6
 Variogram Interpretation
• Variogram interpretation
consists of explaining the
variability over different
distance scales.

• The variogram is a chart of

variance versus distance or
geological variability versus
direction and Euclidean
distance.

• Establish the reasonableness

of the variogram…

13

Some Variogram Terms

• The sill is the variance of the data used for variogram calculation
(1.0 if the data are normal scores)
• The range is the distance at which the variogram reaches the sill
• The nugget effect is the behavior at distances less than the smallest
experimental lag (see next)
– Geological microstructure + measurement error
– Any error in the measurement value or the location assigned to the
measurement translates to a higher nugget effect
– Sparse data may also lead to a higher than expected nugget effect

 Sill

Nugget
Effect
Range
h
14

Lec 03 – page 7
 Comments on the Nugget Effect
• The nugget effect is the variance between pairs that are touching, but not
overlapping

• Could appear different in different directions

• Undefined for distances closer than the

data spacing

• Normally take as isotropic – extrapolate

the V to zero providing closer distances
(that may be needed for H)

15

Trends (1/4)
• Indicates a trend (fining upward, …)
• Could be interpreted as a fractal, fit a power law
function
• Model to the theoretical sill; the data will ensure that
the trend appears in the final model
• May have to explicitly account for the trend in later
simulation/modeling

3.0
Vertical

0.0

-3.0

Horizontal
16

Lec 03 – page 8
 Cyclicity (2/4)

• Cyclicity may be linked to underlying geological periodicity

• Could be due to limited data
• Focus on the nugget effect and a reasonable estimate of the
range

17

Geometric Anisotropy (3/4)

• The range of correlation depends on direction; typically, the
vertical range of correlation is much less than horizontal range
due to larger lateral distance during deposition.
• The shape of variograms can be the same – although the
distance scale is different – this is geometric anisotropy

3.0
Vertical

0.0

-3.0

Horizontal

18

Lec 03 – page 9
 Zonal Anisotropy (4/4)
• Compare the apparent vertical sill with the apparent horizontal sill
• When the vertical variogram reaches a higher sill:
– likely due to additional variance from stratification/layering
• When the vertical variogram reaches a lower sill:
– likely due to an areal trend – areal zoning

Variability between strata 3.0

Vertical
0.0
Variability within strata

-3.0

Horizontal 19

Some Variograms

• Superposition of the four

interpretation principles
– Trend
– Cyclicity
– Geometric anisotropy
– Zonal anisotropy
• Very few variograms are
textbook variograms

20

Lec 03 – page 10
 Review of Main Points

• Variograms are very important in geostatistical studies

• Measure of geological distance
• Initial coordinate and data transformation may be required.
• Interpretation Principles:
– Trend
– Cyclicity
– Geometric Anisotropy
– Zonal Anisotropy
• Short scale structure is most important
• Vertical direction is typically well informed
– can have artifacts due to spacing of core data
– handle vertical trends and areal variations
• Some directions may not be well informed
– take from analog data, outcrops, geological interpretations
– typical horizontal vertical anisotropy ratios

21

Lec 03 – page 11
Geostatistics for Mineral Resources Uncertainty Assessment

Variogram Modeling

• Reasons for Variogram Modeling

• Variogram Modeling
• Inference in Presence of Sparse Data
• Semi-Automatic Fitting

Reasons for Modeling

1. Need to know the variogram for all distances and directions –
not just the ones calculated

2. Incorporate additional geological knowledge (analogue

information or information on directions of continuity …)

3. The variogram model must be positive definite (a legitimate

measure of distance) for subsequent estimation and simulation

Lec 04 – page 1
 Common Variogram Models
Commonly
No spatial correlation
encountered
Should be a small
variogram shape
component of the
overall variance

Similar to spherical
Implies short scale
but rises more
continuity; parabolic
steeply and reaches
behavior at the origin,
the sill
instead of linear
asymptotically

For periodic For periodic

variables variables, when the
period is not regular

Linear Model of Regionalization

• This is universally used in geostatistics to explain complex patterns of
spatial correlation:
nst
Z  u   m  a0Y0  u   a1Y1  u   a2Y2  u     m   aiYi  u 
i 0
• Where m is the mean and the Ys are independent standard factors:

E Yi  u   0 

  
E Yi  u   1  i, j  i, u, u '
2


E Yi  u  Y j  u '  0 

• Each Y factor has its own variogram with different shape, anisotropy,…
i  h  , i  0,..., nst

• By convention, the 0th factor has a nugget effect

4

Lec 04 – page 2
 LMR – Nested Structures
• The variogram of the combined variable depends on the contributions of
the factors:
nst
Z  u   m   aiYi  u 
i 0
nst nst
 Z  h    ai ai i  h    ci i  h 
i 0 i 0

• To model a variogram:
1. Choose nst based on complexity (in practice, no more than four)
2. Choose shape of each structure
3. Choose angles and ranges of each structure
4. Choose contribution of each structure
5. Iteratively adjust as necessary

Nested
Structures

Lec 04 – page 3
 A Complicated Case
Vertical Horz: Major Horz: Minor

0.8 0.8

0.4
0.3
0.1 h
0 8 20 0 400 1000 0 200 1000
Type 2 ax ay az

1.- nugget 0.1

2.- Sph 0.2 8 0 0  h   0.1  0.2  sphav8  0.1  sphav8  0.4  sphav8  0.2  sphav8
ah10 ah10 ah1400 ah1
3.- Sph 0.1 8 0 200 ah 20 ah 2 200 ah 2200 ah 2

4.- Sph 0.4 8 400 200

5.- Sph 0.2 8 9999 9999

Some Variogram Models

• Published variograms
• Characteristic signature for different settings
make it possible to apply with limited data

Lec 04 – page 4
 Inference in Presence of
Sparse Data
• Most often there are inadequate data to infer a reliable horizontal variogram.
• Horizontal wells have not significantly helped with horizontal variograms:
– hard core data and high quality well logs are rarely collected from horizontal wells
– horizontal wells rarely track the stratigraphic coordinate
• At present, we depend on analogue data deemed relevant to the site being
considered such as other, more extensively sampled, reservoirs, geological
process simulation, or outcrop measurements, e.g.,

• A sensitivity analysis must often be conducted

9

In Presence of Sparse Data

• Assemble experimental variogram data: the vertical variogram and the

horizontal variogram lags that can be calculated
• Establish zonal anisotropy: estimate the fraction of variance explained
“within zones” and “between zones”
• Establish presence of vertical / horizontal trends.

Lec 04 – page 5
 Semi-Automatic Fitting
• There is software that mimics the iterative procedure performed by the
modeler, varying the different parameters that define a variogram model.
• The nugget effect, sill contributions, structure types and ranges, are fit to
experimental variogram points in up to three directions simultaneously.
• The user can fix any subset of parameters.

11

McMurray Oil Sands

12

Lec 04 – page 6
 Review of Main Points
• Variogram modeling is one of the most important steps in a
geostatistical study - all spatial inference depends on model of spatial
variability / continuity

• Reasons for variogram modeling:

– need (h) for all distances and directions
– incorporate soft geological information
– must be a licit measure of “distance”

• Variogram modeling principles – linear model of regionalization

– divide total variability into nested structures
– adjust ranges in each direction for each nested structure
– good conceptual model plus trial-and-error

• Inference in presence of sparse data can be a real problem – use

geological understanding and perform sensitivity study

13

Lec 04 – page 7
Geostatistics for Mineral Resources Uncertainty Assessment

Theory of Kriging

• Linear Estimation
• Derivation of Kriging
• Comments on Kriging
• Cross Validation

Context
• The goal is to compute a best estimate at an unsampled location
• Consider the data as differences from their mean values

y  ui   z  ui   m  ui   yi , i  1,..., n
• Statistical inference and a decision of stationarity provides the required
information:

E Y  u   0 


E Y u 
2
 2 
 u  A


E Y  ui Y  u j    C  ui  u j   Cij 

• Covariances are calculated from the variogram model
2

Lec 05 – page 1
 Estimation
• Consider an unsampled location

y *  u 0   z *  u 0   m  u 0   y*  ?
• The only basis to estimate different than the mean is when there are data
indicating that the local area is higher or lower than average

• Give nearby data some influence or weight if deemed relevant:

n
y*   i  yi
i 1
• Weights could be positive or negative depending on relationship
between unsampled location and data event
• Considering quadratic or higher order terms increases inference and
does not lead to improved estimates

Error Variance
• We require a measure of the goodness of an estimate to get the best
• Minimizing the squared error is reasonable


 E2  E  y *  y 
2

• The true value is not precisely known, but we can still calculate the
error variance in expected value:
• This error variance can be calculated using the variogram / covariance
• The mean or expected value (the local conditional one) is always the
estimate that minimizes the squared error criterion
• The estimate that minimizes this error variance by construction will be
called kriging in geostatistics

Lec 05 – page 2
 Recall Variogram to Covariance

• For simplicity we will work with the residuals of the data.

C(0) = sill
• The Variogram is defined by: (h)
C(h)
• The Covariance is defined by:

• These are linked by:

• So:

Kriging

• Consider a linear estimator:

• Expand the error variance:

• Take the partial derivatives and set them equal to zero to find the
minimum:

• This is called simple kriging, although there is nothing particularly simple

Lec 05 – page 3
 Kriging System of Equations

• The three equations:

or

Summary of Kriging

• Kriging uses the variogram to understand the variability of the data

over a distance.
• This knowledge allows kriging to calculate weights that minimize the
error variance.
• It also accounts for other factors that could affect the weights:

– closeness to the location being estimated

– redundancy between the data values
– anisotropic continuity (preferential direction)
– magnitude of continuity / variability

Lec 05 – page 4
 Cross Validation with Kriging
• Cross validation deletes one input data at a time and re-estimates
that point using the remaining data.
• This process is as difficult as the actual estimation of the unsampled
locations – remove the entire well/drillhole
• Sometimes the result is a pessimistic

• The main use of these tools is to detect blunders and problems in the
input data before bad estimates are made

Analyzing Re-estimation Scores

• A scatterplot of the true values versus

the estimated values is perhaps the
most powerful diagnostic tool.
• The distribution of the errors should be
symmetric, centered on a zero mean,
with a minimum spread.
• Mean error and mean squared error
(MSE) are useful summaries.
• The plot of the error versus the
estimated value should be centered
around the zero-error line, a property
called “conditional unbiasedness”
• The errors for a given procedure should
be independent from each other. This
can be checked by contouring or
posting the error values.

10

Lec 05 – page 5
 Continuous Variable Display
• Continuous variables are relatively straightforward – with enough data
• Display points and summary statistics

11

Categorical Variable Display

• Categorical variables (indicators) look different
• Display points and summary statistics

12

Lec 05 – page 6
 Review of Main Points

• Kriging is referred to as the Best Linear Unbiased Estimate (BLUE)

• The weights account for the closeness of the data and the redundancy
between them
• The estimate is unbiased
• Simple Kriging (SK) – has no constraints, but the mean must be known
• Kriging will produce estimates equal to the mean if there are no data
• Cross validation:
– Test how accurately data can be estimated from surrounding data
– Use an assortment of plots and figures to check the results
– Jackknife reestimates a separate set of data (a truly blind test)

13

Lec 05 – page 7
Geostatistics for Mineral Resources Uncertainty Assessment

Practice of Kriging

• Recall
• Ordinary Kriging
• Universal Kriging
• Smoothing

Simple Kriging
• Local estimation n
y *  u 0   z *  u 0   m  u 0   y*   i  yi
i 1


 E2  E  y *  y 
2

n n n
  2  2 i Ci 0   i  j Cij
i 1 i 1 j 1

n
• Linear estimate
• Minimum error variance   C
j 1
j ij  Ci 0 , i  1,..., n
• Nice system of equations
n
 2
SK     i Ci 0
2

i 1
2

Lec 06 – page 1
 Types of Kriging
• Kriging is a procedure for constructing a minimum error variance
linear estimate at a location where the true value is unknown
n
z *  u 0   m   i   z  ui   m 
i 1

• The main controls on the kriging weights are:

– closeness of the data to the location being estimated
– redundancy between the data
– the variogram
• Simple Kriging (SK) does not constrain the weights and works with
the residual from the mean, so the mean has to be known

• Two common variations:

– Constrain the weights to filter a form of the mean
– Transform the data prior to estimation

Kriging with Locally Varying Mean

• Conventional SK assumes a constant mean m everywhere
• The trend could be modeled (e.g. using a moving average technique), so
at every location the value m(u) is known.
• Kriging with a locally varying mean → kriging the residuals.
n
z *  u 0   m  u 0    i   z  ui   m  ui  
i 1
• Need a covariance model for the residuals
– Difficult inference problem
– Consider the variogram of the data
– Perhaps take zones or areas where mean is relatively constant

Lec 06 – page 2
 Ordinary Kriging
• Assume mean is constant but unknown:

would not need m if

• Estimation variance:

to be minimized under the condition that

Ordinary Kriging

• Lagrange Formalism

• Ordinary Kriging system:

• Ordinary kriging variance:

Lec 06 – page 3
 Types of Kriging
• Kriging with a trend model (KT) considers that m is unknown and that
it has a more complex trend of known shape but unknown parameters.
L
m(u)   al f l (u)
l 0
where m(u) is the local mean, a1, l = 0...,L are unknown coefficients of
the trend model, and fl(u) are low order monomials of the coordinates
including( x , y , z , xx , yy , zz, xy, xz , and, yz )
• Kriging with an external drift is an extension of KT. Considers a single
trend function f1(u) defined at each location from some external
(secondary) variable.

• Cokriging is for kriging with different types of data.

• Indicator Kriging has a different goal  to build a conditional cdf rather
than derive an estimate.

Kriging with
External
Drift:

any
Secondary
Variable can
be Used

Lec 06 – page 4
 Kriging
• All versions of kriging are elaborations on the basic linear regression algorithm
and corresponding estimator: n
*
[ Z SK (u)  m(u)]    (u)[ Z (u )  m(u )]
 1
where Z(u) is the RV model at location u, the u 's are the n data locations,
m(u) = E{Z(u) is the location-dependent expected value of RV Z(u), and
ZSK*(u) is the linear regression estimator, also called the “simple kriging” (SK)
estimator.
• The SK weights  (u) are given by the normal equations:
n



 (u)C (u  , u )  C (u, u ) ,
1
  1,...., n

• Some Remarks:
– there are many types of kriging where specific constraints or methods of application
are considered
– the weights  (u) account for (1) the proximity of the data to the location being
estimated and (2) the clustering of the data
– traditionally used for mapping
– modern use is in the construction of local distributions of uncertainty for stochastic
simulation algorithms
9

Variance of Estimates
• Let’s calculate how smooth the estimates are:


Var  Y *  u   E  Y *  u   m 
2


 E  Y * 
2
  n n 
 E   i  jYiY j 
 i 1 j 1 

  i  j E  YiY j    i  j Ci , j
n n n n

i 1 j 1 i 1 j 1
n
 n

  SK
2
 2 iCi ,0   2   SK
2
 2   2   iCi ,0    2
i 1  i 1 
    SK
2 2

10

Lec 06 – page 5
 Variance and Smoothing

• Kriging is locally accurate and smooth, appropriate for visualizing trends,

inappropriate for process evaluation where extreme values are important.
• The simple and ordinary kriging variances are:

• However, these variances are too low and the degree of smoothing is
equal to the missing variance.
• The difference between the global, C(0)=2, and kriged variance is the
missing variance.

11

Review of Main Points

• Kriging is referred to as the Best Linear Unbiased Estimate (BLUE)

• The weights account for the closeness of the data and the redundancy
between them
• Ordinary Kriging (OK) – the sum of the weights is constrained to equal
1.0, which amounts to locally estimate the mean value
• Kriging will produce smooth estimates and the degree of smoothing is
known
• Many variants developed over the years, but Simple Kriging (SK) is the
most important:
o Equivalent to the normal equations in an MG context
o Minimum error variance

12

Lec 06 – page 6
Geostatistics for Mineral Resources Uncertainty Assessment

Principles of Simulation

• Preliminary remarks
• Motivation for simulation
• Setup

Why Simulation?
• Geological formations are heterogeneous at all scales
• Access to formations is costly
• Direct measurements are relatively widely spaced

• Estimation techniques with any rational measure of goodness (mean

squared error…) provide predictions that are smooth

– Estimates at nearby locations

are necessarily similar Unsampled
locations

• Resource and reserve calculations must consider:

– Influence of heterogeneity and variability on recovery Simulation is
– Uncertainty in production and economic forecasts Required
– Optimization in presence of variability and uncertainty

Lec 07 – page 1
 Some Initial Remarks
• Experimental mathematics is well established
• Drop a needle on floor of hardwood strips to calculate …
• (Re)invented many times in last centuries. On the computer:
• The name Monte Carlo Simulation (MCS) was coined by von Neumann
and Ulam as part of the secret research on the Manhattan Project
• Monte Carlo Simulation (MCS) aims to transfer uncertainty:

• The input variables are uncertain (with a model of that uncertainty)

• The transfer function must be known
• In most cases, there is no simple analytical approach to go from input
variable uncertainty to response variable uncertainty
• The dimensionality of the problem may be very high

Initial Example
• What is the sum of the numbers showing on the face of three fair cubic
dice if they were thrown randomly?
• The input uncertainty and transfer function are known
• Simulation proceeds by:
– Draw three random numbers and associated outcomes from input CDF
– Process through the transfer function
– Repeat some number of times (a hundred?)

r3
r1

r2

• The probability values on a CDF are always uniform between 0 and 1

Lec 07 – page 2
 Space of Uncertainty
• The space of uncertainty for the initial example of three dice is small

6 x 6 x 6 = 63 = 216
could be solved exactly

• In practice, the space of uncertainty is combinatorial and intractable

– Many locations (10s of millions)
– Many variables (10s)
– Variables that vary continuously over orders of magnitude
• Space of uncertainty for geostatistical problems > 1010,000,000,000
There are 1030 g in earth and 1055 g in universe.
• Cannot process all combinations. No need since we just want expected
response and some reasonable summary measures.

More Initial Remarks

• Many other words/terms used: Monte Carlo, sampling, simulation,
computer simulation, stochastic simulation, variables, parameters,
outcomes, samples, realizations, outputs, responses,…
• For our purposes in geostatistics (arbitrary):
– Simulation is the process of generating realizations of input variables that will
be considered in a transfer function.
– Think of Estimation as the process of generating estimates for resources
– Think of Flow simulation processing geological realizations for flow response
– Simulation – process of simulating
– Realization – multivariate outcome of simulation (index by l=1,…,L)
• Univariate simulation is commonly achieved by:
– Draw random number (r) uniformly between 0 and 1
– Associate to quantile: z=F-1(r)
• Essential to have representative distribution
• Categorical variables must be ordered, but ordering does not matter
• Essential to have good pseudo random numbers; well established

Lec 07 – page 3
 Uniform Random Numbers
• An long sequence of random numbers is required for simulation.
• 1927: Tippett published a table of 40,000 digits from census reports.
• 1939: Kendall and Babington-Smith create a mechanical device
(machine) to generate random numbers.
• 1946: Von Neumann proposed the middle square method.
• 1948: Lehmer introduced the linear congruential method.
• 1955: RAND Corporation published digits from electronic noise.
• 1965: MacLaren and Marsaglia combined two congruential generators.
• 1989: Wikramaratna proposed the additive congruential method.

Problem Setup
• Geostatistical Simulation: Zkl u , k  1,...,K;l  1,..., L,u  A

• Realizations represent our state of incomplete knowledge:

Lec 07 – page 4
 Scenarios
• Realizations are generated one-for-one and not branched

http://www.waldeneffect.org
http://www.rhynelandscape.com

• Sometimes early in project evaluation Reservoir Level

there are distinct scenarios and

F 0.2 M 0.8 Depositional System
alternative conceptual models
that could be considered Y 0.3 N 0.7 Y 0.3 N 0.7 Fault Seal
(two to four)
• These may be handled I II III I II III IV V IV V Facies
0.4 0.5 0.1 0.4 0.5 0.1 0.4 0.6 0.4 0.6
by scenarios with
Histograms of
branching Uncertainty in
Reservoir Response

Steps in Geostatistical Simulation

1. Model Setup
– What are we going to model, steps, algorithms
2. Prior Parameter Uncertainty
– What are the base case parameters and uncertainty
3. Data Uncertainty
– Fill missing values and sample data error
4. Simulate Realizations
– Apply simulation engine with specified workflow, parameters and data
5. Process in Transfer Function
– Calculate what we are interested in (resources, reserves…)
6. Report Uncertainty and Sensitivity
– Show base case plus uncertainty, evaluate where uncertainty coming from

Lec 07 – page 5
 1. Model Setup
• Specify hierarchical workflow for how a realization of all variables will be
assembled

• Large scale to small scale

• Deterministic to stochastic

• Identify all of the parameters required

for modeling

2. Prior Parameter Uncertainty

• Infer base case parameters and (perhaps correlated) uncertainty in those
parameters

• Mean surface deviations and thickness

– Uncertainty in mean more important than small scale variations

• Facies proportions

• Multivariate spatial bootstrap (combined histogram and correlations)

• Correlation to secondary data

• Variograms

Lec 07 – page 6
 3. Data Uncertainty
• Impute missing values (probably consider decorrelation anyway)

• Consider data uncertainty in the sampling

4. Simulate Realizations
• Apply geostatistical workflow in scripted fashion with specified
parameters and data

• Clip and assemble models as required

• Massively parallel across realizations; only one realization generated at

a time.

• Specify vector of random number seeds indexed to the realization

number, variable number,…

Lec 07 – page 7
 5. Process in Transfer Function
• Compute anything we want on all realizations

• Use a simple transfer function if required and check selected realizations

on full physics transfer function

• Uncertainty and P50 known after processing all realizations

6. Report Uncertainty and Sensitivity

• Uncertainty is straightforward

• Sensitivity calculation is more involved:

– Calculation of scalar posterior parameters
– Fitting a response surface or performing a “leave one out” fitting

Lec 07 – page 8
 Merging of RT and Grade
• Normally take first
RT with the first grade
realization and so on
• A simple template (cookie
cutter) is often used
• The results could be merged
with soft boundaries
– Merge results that share data
across the boundary

• Should merge randomly

– Worst case scenario could
be assembled by low real.

How Many Realizations?

• Many
• Until the results stabilize
• Depends on aspect of uncertainty being targeted
• Consider those realizations that reasonably reproduce all data
• Two views of uncertainty in the uncertainty:
A: Sampling distribution of “P” B: Sampling distribution of “F(p)”

True cdf of performance

Sampling distribution

variable (inaccessible)

Sample quantile from

set of realizations
of F(p)

Sampling distribution of
quantile

• The distribution of B does not depend on the underlying distribution

(distribution in A is problem dependent)

Lec 07 – page 9
 Uncertainty in CDF Value
• The mean and variance of these distributions can be determined
theoretically.
– where the indicator function is 1 if random drawing is less than or equal to
the value F and 0 otherwise. The mean or expected value of F* is the true
underlying cdf value, that is, F.
• The mean of F* is calculated as: L
1
*
F 
L  i(u )
l 1
l

– where the indicator function is 1 if random drawing ul is less than or equal to

the value F and 0 otherwise.
• The variance of F* is calculated as:

 F2 * 
1 2 1
L

 i  E i 2  E
L
 1
i 2  F  F 2 
L

F (1  F )
L
• The distribution shape will approach Gaussian as L increases
• Probability to be inside interval can be easily calculated

Number of Realizations
• The sampling distribution of the probability can be derived
• The following table summarizes the results for common probability
and tolerance values

• Although this seems to indicate that many realizations are required –

the best approach is to consider 100 or 200 realizations
– There are too many subjective decisions in geostatistics and the
realizations are too large to justify creating more
– Professional and computer time better spent improving the model

Lec 07 – page 10
 Final Remarks
• Promote best practice of geological modeling and geostatistics
– Simulation is required
– Principles are well established; details of implementation are not
• Steps in simulation:

• Realizations:

Lec 07 – page 11
Geostatistics for Mineral Resources Uncertainty Assessment

Categorical Variable Modeling

• Preliminary remarks and background

• Context of the course
• General notation and theory

Overview of Techniques
1. Digitization of limits on 2-D sections (reconciled to 3-D) or in 3-D directly
– Always the best approach if possible
– Requires close data relative to scale of variability
2. Semi-deterministic (implicit) modeling with signed distance functions
– Easier than digitization, but less control
– Systematically erode and dilate boundaries
3. Process mimicking models
– Applicable in a few situations
– Nice looking and predictive models
4. Object based models
– Applicable in some situations
– Requires clear object sizes and shapes
5. Cell based models
– Sequential indicator, truncated Gaussian, multiple point statistics
– Statistically control variability between cells
2

Lec 08 – page 1
 General Comments
• Geological sites are modeled hierarchically
– Large scale limits / container (surfaces, digitized limits,…)
– Major heterogeneities (often categorical variables)
– Small scale features (often continuous variables such as mass/volume
fractions and rate constants)
• Categorical variables:
– Based on some geological characteristic of the rock that is large scale
– Normally do not consider more than 5 to 7 at one time
– Categories have different statistics (univariate or spatial)
– Reasonable consistency within and variation between
– Mutually exclusive at the scale of modeling (and exhaustive)
– Spatially consistent/coherent with some degree of predictability
– Must be enough data within each category for inference
– Contact analysis is performed – hard boundaries are preferred

Deterministic Interpretation
• Digitize contacts and reconcile in 3-D
• Not a geostatistical subject, but important

http://s3.amazonaws.com/s3.agoracom.com/public/photos/images/3623/large/Represa%203D.jpg

Lec 08 – page 2
 Boundary Modeling with Distance
Functions
• The limits of each domain must be determined
• Deterministic model with a digitized solid model or surfaces
• Partially stochastic with distance functions

• Distance function is sign dependent

distance to something different:
– Aimed at binary systems
– Modify for uncertainty

• The resulting continuous variable

is very non-stationary

Interpolate the Distance Function

• Interpolate with global kriging or another reasonable estimation technique

• Three dimensions
• Consider locally varying anisotropy if required

Lec 08 – page 3
 Process Mimicking
• Process mimicking or event based modeling is becoming more common
• Very deposit specific
• Some vein type and laterites

Object Based Modeling

• Motivation:
– Visually attractive
– Clean geologic shapes
– Realistic in right circumstances

• Steps:
1. Model large scale features
2. Model small scale features
3. Populate with petrophysical properties

Lec 08 – page 4
 Cell Based Techniques
• Assign a category to a gridded model
• Statistical control with surrounding cells
through variograms or higher order
statistics
• Could work through one or more
continuous variables – truncated (pluri)
Gaussian
• Trend modeling is important for large
scale features

Review of Main Points

• Categorical variables include rock types, facies, geological units,…
• Important to define them carefully to make sense geologically and to
have unique statistical and spatial features
• Model with as much determinism as possible
• Resort to geostatistical modeling as necessary
• Multiple realizations will be necessary to quantify uncertainty

Lec 08 – page 5
Geostatistics for Mineral Resources Uncertainty Assessment

Multivariate Gaussian Distribution

• Multivariate Gaussian
• Congenial properties
• Local conditional distributions

Normal Distribution

• The normal distribution was first introduced by de Moivre in

an article in 1733 (reprinted in the second edition of his
The Doctrine of Chances, 1738) in the context of
approximating certain binomial distributions for large n. His
result was extended by Laplace in his book Analytical
Theory of Probabilities (1812), and is now called the
Theorem of de Moivre-Laplace.

• Laplace used the normal distribution in the analysis of

errors of experiments. The important method of least
squares was introduced by Legendre in 1805. Gauss, who
claimed to have used the method since 1794, justified it
rigorously in 1809 by assuming a normal distribution of the
errors.

Lec 09 – page 1
 Central Limit Theorem (CLT)
• Theorem: the sum of a great number of independent equally distributed (not
necessarily Gaussian) standardized random variables tends to be normally
distributed
• The standard (m=0,=1) normal probability density function: 1 [  z 2 / 2]
f ( z)  e
2

• The cumulative distribution function is not defined analytically, but excellent

numerical approximations exist

More on Gaussian Distribution

• A Z variable is non-standard Gaussian when the standardized variable Y is

standard Gaussian:
z  mZ z  y  Z  mZ
y
Z
• A non standard Gaussian value is easily converted to/from a standard
Gaussian value
• Corollary to CLT: the product of a great number of independent equally
distributed random variables tends to be lognormally distributed:

n
Y   Yi
i 1

X  ln Y
n n
X   ln Yi   X i CLT applies to X's
i 1 i 1

Lec 09 – page 2
 Bivariate Gaussian Distribution

• The Gaussian distribution of 2 variables is defined:

( X , Y )  N (0,1,  X ,Y ) ,
 
 1
x 2

 2  xy  y 2 
1 
 2 1  2  
f X ,Y ( x, y )  e  

2 1   2
• The relationship between the two variables is defined by a single
parameter: the correlation coefficient
• The probability contours are elliptical. Some special parameters:
– Conditional expectation a linear function of conditioning event
Y
E Y X  x  mY   X ,Y ( x  mx )
X
– Conditional variance is independent of conditioning event

Var Y X  x   Y2 (1   X2 ,Y )

Multivariate Gaussian Distribution

• where d is the dimensionality of x. Note that  is a (d x 1) vector and 

is a (d x d) positive definite, symmetric matrix. The expression || is the
determinant of .
• It is also the case that  is the mean of the distribution and  is the
covariance matrix. The i-th element of  expresses the expected value
of the i-th component in the random vector x; similarly, the (i; j)
component of  expresses the expected value of x i x j minus  i  j
• The diagonal elements of  are just the variances of the corresponding
component of x.

Lec 09 – page 3
 Multivariate Gaussian Distribution

• The multivariate (N-variate) Gaussian distribution is defined in matrix

terms (see previous) and remains tractable with some extraordinary
properties:
– All lower order N-k marginal and conditional distributions are Gaussian
– All conditional expectations are linear functions of the conditioning data
E{ X i  X j  x j j  i}    j x j   ( x j  j  i )  [ xi ]SK
j i
– All conditional variances are homoscedastic (data values independent):
2 2
E{ X i   ( x j  j  i )   X j  x j j  i}  E{ X i   ( x j  j  i)  }

• The mean and variance of a conditional Gaussian distribution are given

by Simple Kriging!

Linearity and Homoscedasticity

• Conditional expectations are linear functions of the data

• All linear combinations of normal variables are also normal
• In particular, averages are normal:
1
X V (u) 
V 
V (u)
X ( u ) du is normal

• Conditional variances are data-values-independent

Lec 09 – page 4
 Application of Multivariate Gaussian
Distribution
• Assuming Gaussianity within reasonably defined rock types is a reasonable
approximation
• Let’s make the variable Gaussian and assume multivariate Gaussianity…
• We would only need correlations/covariances/variograms
• We would only need to perform simple (co)kriging to get mean and variance
of all conditional distributions
• All distributions of uncertainty would be Gaussian with mean and variance
specified by:
– Linear combination of data in simple (co)kriging
– Homoscedastic kriging variance
• But….
– Real data distributions are not Gaussian
– Multivariate relationships are not homoscedastic and Gaussian
– Constraints, Heteroscedasticity, and Non-linearity are encountered in practice
• Everything we do not explicitly control in mathematics tends toward a
Gaussian distribution

Gaussian or Normal Score Transform

• Transform any distribution to a Gaussian or normal distribution

• Almost always used for continuous variables
• Steps:
1. Z-variable that follows Fz(z)
2. Y-variable is standard Gaussian (my=0, s2y=1.0) that follows CDF of
standard Gaussian G(y)
3. Transform by “matching” quantiles, y = G-1(F(z)), z

Lec 09 – page 5
 Gaussian Transform
• Univariate transformation guarantees a univariate Gaussian distribution
• There is no guarantee of a multivariate Gaussian distribution :
– Non-linearity is not removed
– Constraints are not removed
• The proportional effect / heteroscedasticity is largely removed by transfromation
and reintroduced by back transformation

Examples of Proportional Effect

Original Original
Units Units

  0.04

Normal Normal
score score

Lec 09 – page 6
 More Examples

  0.98   0.97

Original Original
Units Units

  0.25   0.04

Normal Normal
score score

Quantile Backtransformation
• Quantiles can be back transformed at any time
• Non-standard Gaussian distributions are valid – just back transform non-
standard Gaussian to a possible shape in real data units (Z):

 
z l  FZ1 G  SK G 1 ( p l )  y SK  , l  1,..., L

• Assemble distributions in Z units that are consistent with the multivariate

Gaussian model

• Should not back transform expected values, unless the distribution is

symmetric

Lec 09 – page 7
 Another Picture of Transformation

Review of Main Points

• Prediction of local distributions of uncertainty is a critical problem in
geostatistics
• Bayes law is used one way or the other:
– Directly to get conditional probabilities as ratio of joint probability to marginal
probability of conditioning event
– Indirectly using a Gaussian model for the probabilities and knowing that the ratio
of Gaussian probabilities is also a Gaussian probability
• Data do not follow a Gaussian distribution, but continuous variables are
easily transformed to follow the univariate Gaussian distribution;
multivariate Gaussianity is then assumed
• Categorical variables are often modeled using other means such as object
based modeling or process modeling
• The Gaussian distribution is remarkable in its mathematical tractability – we
only require covariances to infer any order multivariate distribution
– The only practical solution to the curse of dimensionality

Lec 09 – page 8
Geostatistics for Mineral Resources Uncertainty Assessment

Gaussian Simulation

• Preliminary remarks and background

• Context of the course
• General notation and theory

• Estimation is locally
accurate and smooth,
appropriate for visualizing
trends, inappropriate for
engineering calculations
where extreme values are
important, and does not
assess global uncertainty

• Simulation reproduces
histogram, honors spatial
variability (variogram), 
appropriate for flow
simulation, allows an
assessment of uncertainty
with alternative realizations
possible

Lec 10 – page 1
 Preliminary Remarks
• Kriging is smooth – but is entirely based on the data
• Kriging does not reproduce the histogram and variogram
• Simulation is designed to draw realizations that reproduce the data,
histogram and variogram
• The average of multiple simulated realizations is very close to kriging

• Although multiple simulated realizations average to kriging, the

average behavior of multiple realizations may be very different
from that of the kriged model

• The advantages of simulation are:

– Heterogeneity
– Uncertainty
• The price is non-uniqueness

Prerequisites
• Work within statistically homogeneous stationary populations
– Model facies/RTs first
– Chronostratigraphic or unfolded grid framework
• Must have clean data that are positioned correctly with manageable
outliers and error content
• Need to understand special data:
– trends
– production data
– seismic data
• Do not leave data out for checking when building the final model
• Simulation considers grid nodes and not grid blocks:
– Assign a property at cell center at the scale of the data
– Normally paint entire cell based on central value
– Sometimes refine the grid dynamically after simulation

Lec 10 – page 2
 Simulation from High Dimensional
Multivariate Distributions
• MCS or simulation from a univariate distribution is easy

• MCS from a very high dimensional distribution is more complex

– Must ensure the relationship between all values
– Require a multivariate distribution model to sample from
• Sequential simulation can be thought of as a recursive decomposition
of a multivariate Gaussian distribution by Bayes Law

Sequential Conditional Simulation

• Consider N dependent events Ai, i=1,…,N that we want to simulate

• From recursive application of Bayes law:
P( A1 ,..., AN )  P ( AN | A1 ,..., AN 1 )  P ( A1 ,..., AN 1 )
 P ( AN | A1 ,..., AN 1 )  P ( AN 1 | A1 ,..., AN  2 )  P( A1 ,..., AN  2 )

 P ( AN | A1 ,..., AN 1 )  P ( AN 1 | A1 ,..., AN  2 )    P( A2 | A1 )  P( A1 )

• Proceed sequentially to jointly simulate the N events Aj:

– Draw A1 from the marginal P(A1)
– Draw A2 from the conditional P(A2 |A1=a1)
– Draw A3 from the conditional P(A3 | A1=a1 ,A2=a2)
– …
– Draw AN from the conditional P(AN | A1=a1 ,A2=a2 ,…,AN-1=aN-1)

• This is theoretically valid with no approximations/assumptions

6

Lec 10 – page 3
 More Sequential Simulation

• Sequential approach requires:

– Knowledge of the (N-1) conditional distributions
– N simulations with increasing conditioning (larger and larger systems)
– Solution to size problem is to retain only the nearest conditioning values
• Condition to prior data:
– Partition N into n prior and N-n events to be simulated
– Start at n+1 and require N-n conditional probabilities
• Sequential indicator simulation for categorical variables is common
• Sequential Gaussian simulation is widely used:
– Determine each conditional distribution by multivariate Gaussian model
– Virtually all continuous variable simulation is Gaussian based
– ANSI standard algorithm
– Equivalent to turning bands, matrix methods, moving average
methods,…

Sequential Simulation
• Transform variables one-at-a-time to a Gaussian distribution
– Use representative distribution
• Establish the covariance between all pairs of variables / locations
– Variogram of normal scores
• Predict local uncertainty
• Simulation
– Loop over all locations
– Estimate local uncertainty
– Sample uncertainty
• Repeat with different random numbers for multiple realizations
• Back transform to original units

Lec 10 – page 4
 Some Implementation Details
• The sequence of grid nodes to populate:
– Random to avoid potential for artifacts
– Multiple grid to constrain large scale structure

• Assign data to grid nodes for speed

• Keep data at their original locations if working with a small area

• Search to variogram range and keep closest 20-40 data to calculate

local mean and variance

• Could make the mean locally varying

Check Results
• Check in normal units and in original units
• Reproduce data at their locations?
• Reproduce histogram, N(0,1), over many realizations?
• Reproduce variogram?
• Average of many realizations equal to the kriged value?

• Expect statistical fluctuations in the output statistics particularly

when the domain size is small relative to the variogram

10

Lec 10 – page 5
 Different Simulation Algorithms
• There are alternative Gaussian simulation techniques:
– Matrix Approach LU Decomposition of an N x N matrix
– Turning Bands simulate the variable on 1-D lines and combine in 3-D
– Spectral and other Fourier techniques
– Fractal techniques
– Moving Average techniques

• All Gaussian techniques are fundamentally the same since they

assume the multivariate distribution is Gaussian
• Sequential Gaussian Simulation widely used and recommended
because it is theoretically sound and flexible in its implementation

11

Review of Main Points

• Continuous variable modeling:
– Work within homogeneous stationary subsets
– Grid system that facilitates data integration
• Estimation is not appropriate for most engineering applications
• Simulation is useful for quantifying uncertainty on a response
variable (reproduce spatial variability)
• Sequential simulation decomposes the multivariate distribution
into a sequence of univariate conditional distributions
• Gaussian simulation is used because of its “nice” properties
– All conditional distributions are Gaussian in shape
– Conditional mean and variance are given by
normal equations (called simple kriging
in geostatistics)

12

Lec 10 – page 6
Geostatistics for Mineral Resources Uncertainty Assessment

Practice of Simulation

• Implementation
• Checking
• Programs

Steps in Sequential Gaussian Simulation

• transform data to “normal space”
• establish grid network and coordinate system (Zrel-space)
• assign data to the nearest grid node (take the closest of multiple data assigned
to the same node)
• determine a random path through all of the grid nodes
– find nearby data and previously simulated grid nodes
– construct the conditional distribution by kriging
– draw simulated value from conditional distribution
• check results
– honor data?
– honor histogram: N(0,1) - standard normal with a mean of zero and a variance of one?
– honor variogram?
– honor concept of geology?
• back transform

Lec 11 – page 1
 Normal Scores Transformation

“graphical” one-to-one (rank preserving) transform

Cumulative Distribution Cumulative Distribution

Cumulative Frequency

Grade Normal Scores

Frequency

Transformation to/from Normal Space

• Declustering weights are necessary

• Use “global” transformation table for small-area or cross-sectional models 
there may be too few data in the area of interest
• “Tail” options:
– typically simulate the property at many more grid nodes than there are data  expect
higher values than observed in the data (and lower).
– almost always a linear interpolation to user-defined minimum and maximum values is
acceptable
– can account for skewness in the upper tail with a hyperbolic model

5.0

1.0

F , ( z )  1  ,  1
z
F(zk)

zk Z variable

Lec 11 – page 2
Two-Part Search or Assign Data to
Nodes

Two-Part?
• search for previously simulated nodes and then original data in two steps 
then treat the same when constructing the conditional distribution
• honor the data at their locations even if they can not be seen in the final model
• necessary for cross-sectional or small-area models
Assign Data to Grid Nodes:
• explicitly honor data - data values will appear in final 3-D model
• improves the CPU speed of the algorithm: searching for previously simulated
nodes and original data is accomplished in one step

Multiple Grid Concept

• We typically limit ourselves to the nearest 12-48 grid nodes

• As the sequential simulation proceeds (fine grid) only the close samples will be
used
• Variogram reproduction can be poor, especially if the variogram range is large
with respect to the grid node spacing
• Multiple grid concept: simulate a coarse grid first and then refine one or more
times
• Searching for data is more complicated:
– two-part search at coarse levels
– perhaps relocate to nearest fine grid node first (unless a two-part is used at the finest
level)
• A number of implementation considerations

Lec 11 – page 3
 Number of Data to Consider
• Reasons for more:
– theoretically better
– more accurate estimate of the conditional mean and variance
– better reproduction of the variogram
• Reasons for less:
– CPU time is proportional to N3
– memory requirements proportional to N2
– negative weights are commonly encountered when data are screened
– using fewer data places less emphasis on the assumption of stationarity
• So, choose between 12 to 48 depending on:
– 2-D versus 3-D
– range of variogram relative to grid node spacing
– CPU time restrictions

Type of Kriging
• Simple Kriging (SK) is equivalent to normal equations and is theoretically
correct:
  i  z ( ui )  1   i   m global
n n
*
m SK
i 1
 i 1 

• Simple cokriging works well

• Locally varying mean can work, but variance can be too high – reduce
the sill if needed
• Ordinary Kriging (OK) and other constrained varieties used with care
because of a too-high variance

Lec 11 – page 4
 Ergodic Fluctuations
• A statistical domain is ergodic when it is large with respect to the scale of
variability
• Expect some statistical fluctuations in the input statistics when the
domain is not ergodic

Check the Results

• Basics:
– honor the data?
– reproduce declustered distribution reasonably well?
– reproduce the variogram reasonably well?
• “Validation”:
– honor the geological interpretation?
– require a lithology/rock-type model?
– require a trend model?

Lec 11 – page 5
 Review of Main Points
• Kriged estimates are too smooth and inappropriate for most engineering
applications
• Simulation corrects for this smoothness and ensures that the variogram /
covariance is honored
• There are many different simulation algorithms → sequential Gaussian is simple
and most widely used
• Use Gaussian / normal distribution for a consistent multivariate distribution

Lec 11 – page 6
Geostatistics for Mineral Resources Uncertainty Assessment

Uncertainty and Sensitivity Analysis

• Preliminary remarks and background

• Context of the course
• General notation and theory

Quantifying Importance
• Impact of a variable on the local uncertainty.
• Remove the variable from the system, and reassess kriging variance
• Importance – the deviation in kriging variance observed when that
variable is removed from the system.
 K2 ,i   K2
Importancei 
 K2
i variable(s) removed
 K2 ,i kriging variance with variable i removed
 K2 kriging variance when all variables are considered.

• Full combinatorial analysis

– i represents a combination of removed variables
– Remove from 1 to N-2 variables
• N – total number of variables

Lec 12 – page 1
 Importance Measures

Context and Motivation

Response
Variable

Transfer f(y)
function

P10 E(y) P90 y

Adapted from Clayton (2015). CCG Annual Meeting 2015:

02_Parameter_Uncertainty

We need a tool to visualize, summarize, analyse and present
uncertainty and sensitivity

Lec 12 – page 2
 First Goal

Develop a tool to display uncertainty sensitivity analysis

Ideal Characteristics
Easy to understand (essential information)
• Simple
Easy to use (“friendly”)

• Informative

What kind of graphs?

Tornado Chart

Uncertainty in  Contribution? Uncertainty in 

input variables  model response

Lec 12 – page 3
Contribution to Uncertainty in Response?

R2=0.92 R2=0.23

Lec 12 – page 4
Geostatistics for Mineral Resources Uncertainty Assessment

Overview of Multivariate Geostatistics

• Taxonomy of techniques
• Data exploration and preparation
• Model building

Some Important Workflows

I. Data Preparation
a. Transform Compositional Data
b. Transform to Normal Scores
c. Summarize Multivariate Relationships
d. Aggregate Data Variables An attempt to
e. Cluster Data Observations organize what we
commonly do in
f. Fit Response Surface
multivariate
geostatistics
II. Spatial Prediction
a. Estimate with Unequally Sampled Data
b. Simulate with Colocated Cokriging
c. Impute Missing Data
d. Simulate with Linear Decorrelation
e. Simulate with Nonlinear Decorrelation
f. Simulate with Stepwise Conditional Transform
2

Lec 13 – page 1
 I.a Transform Compositional Data
• Modeled variables commonly have
constraints which must be obeyed
– Compositional components must not sum to
greater than 100%
– Net measurements must not exceed associated
gross (acid soluble versus total)
• Consider a set of variables that are a
composition (perhaps with a filler variable)
with last variable as reference:
K
zk , k  1,..., K with zk  0 and z
k 1
k 1

• Consider ratios, perhaps with log:

zk
xk  , k  1,..., K
zK

Program: logratio
3

I.b Transform to Normal Scores

• Multivariate techniques require data to be standardized
• Normal score transform is non-linear, but that is fine for simulation
• Often applied in a chained transformation (all reversible)
• There are weak checks for multivariate Gaussianity after univariate
• Must transform many at once (not one at a time)
• Must have representative distribution to transform from/to

Program: unscore (multiple variables and rocktypes) 4

Lec 13 – page 2
I.c Summarize Multivariate Relationships
• Cross plots are a pale shadow of full multivariate space, but…
• Correlation matrix is interesting – may consider optimal ordering
• Multidimensional scaling may help group or screen the variables
• MDS could be followed by clustering

Program: corrmat (correlation matrix, MDS, regression) 5

I.d Aggregate Data Variables

• Merge data variables while maintaining the information from each
variable (relative to a particular variable)
ns ns
ySS     x 

  
1
x , x   x , y ,   1,..., ns
 1
ns
 SS  

  
1
x , y
LEVEL 2:
A B C D
Cu(wt%) Merged_1 Merged_4 Merged_7
U3O8(ppm) Merged_2 Merged_5 Merged_8
SG Merged_3 Merged_6 Merged_9
4 final variables K:Al Merged_10
Ag(ppm) Merged_11
Au(ppm) Merged_12
Badj%S Merged_13
Merged_14

• Simplifies geostatistical modeling Merged_15

Merged_16

• Avoids overfitting
• Influence of each can be determined by sampling

Program: supersec
6

Lec 13 – page 3
 I.e Cluster Data Observations
• Group observations that are close
in multivariate space

• Main algorithms:
– Hierarchical
– K-means
– Gaussian Mixture Model

• Review in dendrogram and look at

2-D projections

Program: cluster
• Several clustering
algorithms
• Dendrogram plot (for hier.)
• Summary plots
7

I.f Fit Response Surface

• Fit a response variable (Y) by some number of predictor variables (Xs)
– Linear NSV NSV NSV

– Quadratic Yˆ  b0   bi X i   ckl X k X l
i 1 k 1 l  k
– ACE

Programs: ace and sabor

8

Lec 13 – page 4
II.a Estimate with Unequally Sampled Data
• Many situations where data are never (or rarely) equally sampled
– Diamond drill holes vs. reverse circulation.
– Blast hole samples vs. exploration drill hole samples.
– Legacy data

• Cokriging with conventional

linear model of cogregionalization

• Key idea is to use cross covariance

Program: cokb3d (new version) 9

II.b Simulate with Colocated Cokriging

• Colocated Cokriging a powerful approach in simulation:
– The colocated secondary variable is considered
– Cross covariance CZY(h) is assumed to be a linear scaling of CZZ(h)
• The collocated value is surely the most important and screens the
influence of multiple secondary data
• Intrinsic colocated cokriging corrects for variance inflation and other
artifacts – highly recommended

Program: usgsim (new version) 10

Lec 13 – page 5
 II.c Impute Missing Data
• Multivariate transforms may only be executed on equally sampled
(homotopic) observations
• Multiple imputation is the best approach in most circumstances

• Each imputed data realization conditions its own geostatistical realization

• Two different approaches for this developed by CCG
– Gibbs sampler approach by Ryan Barnett
– Gaussian mixture model approach by Diogo Silva
Programs: impute and gmm_impute 11

II.d Simulate with Linear Decorrelation

• Linear decorrelation well
established:
– PCA – only way to decorrelate
– Sphere – return variances to one
– Rotate again
• For spatial structures (MAF)
• Back to original basis (R)
• Dimension reduction possible
• May lead to deeper understanding

Program: decorrelate 12

Lec 13 – page 6
II.e Simulate with Nonlinear Decorrelation
• Transform the multiple variables with a multivariate nonlinear transform
to decorrelate and make multivariate Gaussian
• Projection Pursuit Multivariate Transform (PPMT) now widely used

Programs: ppmt and ppmt_b 13

II.f Simulate with

Stepwise Conditional
Transform
• Transform in stepwise fashion
Y1  G 1  F1 ( Z1 ) 
Y2  G 1  F2|1 ( Z 2 | Y1 ) 

YK  G 1  FK |1,..., K 1 ( Z K | Y1,..., K 1 ) 
• Limited number of variables, but works
well with secondary data
• Implementation developments including
KDE, GMM,…
Programs: sctrans and gmm_sct 14

Lec 13 – page 7
 Comments on Workflows
• These workflows are stable and
well established
• CCG has tools for all of these
workflows

• There are other workflows and

alternative algorithms
• Understanding the theory and
fundamentals is important

• Simulation algorithms compete:

– Choose simplest considering
multivariate complexity of data
– They are related theoretically

15

Lec 13 – page 8
Geostatistics for Mineral Resources Uncertainty Assessment

Multivariate Simulation with

Decorrelation
• Principal component analysis
• Sphering
• MAF
• Applications

Background
• PCA is a classic dimension reduction and
decorrelation technique that was
developed by Pearson (1901) and
Hotelling (1933)
– Adapted to geostatistics by Joreskog et al
(1976) , as well as Davis and Greenes
(1986)

• Transforms a correlated multivariate

distribution into orthogonal linear
combinations of the original variables
– Variables may be modeled independently

• PCA often used for exploratory analysis

– Transformed variables explain a
descending amount of variability
• May consider dimension reduction 2 2

Lec 14 – page 1
 Linear Decorrelation Workflow
• An elegant workflow for multiGaussian data
– Covariance matrix fully explains the multivariate distribution
• Decorrelation leads to independence

PCA Theory
• The primary objective is often to transform correlated variables into
orthogonal principal components.

• Let K centered RVs (residuals) be denoted as Zi where i=1,…, K

– Consider standardizing or normal score the data matrix to center

Z : z i ,   1,..., n, i  1,..., K
• The covariance matrix of the data is given as ΣZ  1 n ZT Z

• Define Y as the transformed data matrix, such that Yi , i = 1,…, K are

uncorrelated with each other, possess zero mean, and are the linear
combination of Z
C11  0 
 
Y = ZA, where A A  I T
ΣY  1 n Y Y    
T
 
 0  zK K 
  4

Lec 14 – page 2
 Data Exploration with PCA
• PCA is often used for applications beyond decorrelation, such as
sensitivity analysis and dimension reduction
Σ Z  VDV T
Orthogonal eigenvectors Diagonal Eigenvalues
 v11  v1 K   d11  0 
   
V     D    
vK 1  vK K   0  dK K 
   

• Proportion of variance explained by Yi  di / tr(D)

• Importance of Yi to Z j indicated by ρ'i j

– Proportional to the correlation between Zi and Yk
– Also calculated as the product of V and D

Nickel Laterite Loadings

• Loadings explain the

nature of the rotation
– Quantify how each of
the original variables
are loaded onto the
transformed variables
– Or, how the transformed
variables explain the
original variables
• The dimension reduction
property of PCA is
evident in the resultant
loadings
– Can be derived as an
optimization problem
that maximizes
6

Lec 14 – page 3
 Nickel Laterite Dimension Reduction
• The variability of the first four prinicipal
components of the Nickel laterite
system according to d i / tr( D)

• Consider back-transforming the

transformed data using only the first
two principal components
– Scatter plots compare data of the
four variables (X-axis), vs. their
back-transformation (Y-axis)

Scree Plots
• Help decide # of PCs to use
– Choose up to and including point on an “elbow”
– Sensitive to sample size

• Scree plot usefulness

– Check random subsets
Raw Cumulative Standardized Cumulative
1 55.0203 55.0203 52.7686 52.7686
2 43.8491 98.8693 22.8726 75.6411
3 0.7601 99.6294 13.7272 89.3683
4 0.2707 99.9001 8.1585 97.5269
5 0.0999 99.9999 2.3306 99.8575
6 0.0001 100.0000 0.1425 100.0000

Lec 14 – page 4
 Data Sphereing
• Sphereing is closely related to PCA
– PCA: , where
– Sphereing (or simply, Sphere) refers here to the
standardization of the rotated variables according
to:
– Sphere-R (reverse projection) refers here to the
standardization of the rotated variables before
projecting the variables back onto the original basis:

• Sphereing may marginally simplify subsequent

analysis relative to PCA
– Sphere-R is more noteworthy since it minimizes
variable mixing

Sphere and Sphere-R Loadings

• The Sphere transformation is very similar
to PCA in that it maximizes the total
variance that is explained by each rotated
component
– Maximizes the absolute value of
values for all i

• The Sphere-R transformation maximizes

the variance that each rotated explains
about the associated original variable
minimizing the variance it explains about
the other variables
– Maximizes the absolute value of
values for i = 1 while minimizing the
absolute value of for i ≠ 1

10

Lec 14 – page 5
 MAF Background

• First introduced by Switzer and Green (1984) in the field of spatial

sensing and popularized in geostatistics by Dimitrakopoulos et.al.
(2000)

• Extension of PCA – usually interpreted as applying PCA twice:

1. Lag h = 0  exactly the same as PCA
2. Lag h  attempts to destructure the remaining spatial correlation

• Results in random variables that are spatially uncorrelated at a 0 and

non-zero lag distances
– Leads to better reproduction of cross-variograms in simulated realizations

• The following lecture will outline the ‘data-driven’ MAF approach

– Dimitrakopoulos et.al. (2000) also proposed a ‘model-driven’ MAF approach
that requires the prior fitting of an LMC model to inform the spatial
decorrelation 11

Illustration of Rotation and Cross

Correlation

12

Lec 14 – page 6
 Transformed Loadings
• Loadings of Sphere and Sphere-R
transformation explain the nature of those
techniques
– Relate to variance of the h = 0 multivariate
system
• Loadings of MAF appear more unorganized
– Relate to variance of the h > 0 multivariate
system
• MAF rotation
is more easily
understood
after viewing
the
transformed
variograms

13

Review of Main Points

• PCA is a classic technique that serves many functions
– Exploratory analysis
– Dimension reduction
– Decorrelation
• Is often used as a preprocessor for independent geostatistical modeling.
Use with caution:
• Decorrelation does not mean independent unless the data are also
multivariate Gaussian
• Principal components will be mixtures of original variables, which could
create challenges for reproducing properties such as the variogram
following simulation and back-transformation
– Sphere-R minimizes the mixing of variables
• Independence at h = 0 does not mean independence at h > 0
– MAF rotation removes spatial correlation

14

Lec 14 – page 7
Geostatistics for Mineral Resources Uncertainty Assessment

Principles of Geometallurgical Modeling

• Averaging schema
• Linearizing transforms
• Multiscale simulation
• Measurement error
• Case study

Geometallurgy
• One perspective

• Some focus on mineralogy

• Some focus on metallurgy

Lec 15 – page 1
 Geometallurgical Data

Geometallurgical
Geostatistics
• Proposed approach for
challenging data is data
transformation followed by
either:

1. A correlated multivariate
simulation workflow

2. Multivariate and spatial

decorrelation transforms
followed by an independent
simulation workflow

• Transforms are reversed after

modeling

Lec 15 – page 2
 Data Transformation Methods

What Do We Mean by Non-linear in?

• In this context, we are referring to variables that do not average linearly
when mixed
– The Bond mill work index (BMWi) is a classic example
• The most widely used test for grindability in mineral processing mills is
BMWi
– Gives the work input required to grind a particle from a very large size to a
point where 80% of the product passes100 microns
• BMWi tends to be
biased towards the
more energy intensive
ore
– Vivians in the right
figure is relatively soft
material

Lec 15 – page 3
 Bitumen Recovery
• Another variable exhibiting non-linear behaviour is bitumen recovery
• Increasing the concentration of fines for oilsands flotation sharply
decreases bitumen recovery in the froth
• Response is percolation-like: little effect of fines on bitumen recovery up
to 12% fines, then recovery drops substantially

Sanford, E. C. (1983). Processibility of athabasca oil sand: Interrelationship between oil sand fine solids, process aids, mechanical 7
energy and oil sand age after mining. The Canadian Journal of Chemical Engineering, 61(4), 554-567.

Mixing is Usually Unavoidable

• Even when batch fed, a large amount of mixing will occur
– Two extreme feed scenarios below, both of which lead to mixing

• For geostatisticians, this means we cannot just ignore the nonlinearity in

metallurgical variables, or assume that by batch feeding we are
sidestepping antagonistic blending behaviour

Lec 15 – page 4
 Methods to Infer Nonlinearity
• Ideally use direct
measurement of blends of
known ore types to infer
nonlinearity

• With multiscale
measurements, nonlinear
regression may be used,
however error will bias the
regression

• Given multiscale correlated

measurements, use
semiparametric Bayesian
updating with nonlinear
regression

Nonlinear Inference with Multiscale Data

• Multiscale measurements occur in a number of cases:
– Subsampling ore, testing, blending and re-testing
– Small scale correspondence measurements

• Apply nonlinear regression to infer the nonlinearity

Lec 15 – page 5
 Impact of Measurement Error
• Measurement error, particularly in the case of small scale
correspondence measurements, is expected
• Results in regression attenuation, a reduction in signal
strength due to measurement error (Spearman, 1904)
• Error in measurements
may lead to bias in
models, power of 3.5 fit
instead of 5.5 for the
case shown, although it
appears unbiased

Attenuation of Power Law Models

• For true linear models with normal, homoscedastic, unbiased error on
the observed, large scale parameter, no bias is observed
• No bias observed for this case with nonlinear models as well

Lec 15 – page 6
 Attenuation of Power Law Models
• If measurement error is present on the small scale measurements, then
the fit will be biased for substantial amounts of error, even in a linear
model
• This is expected behavior, we could consider disattenuation

Attenuation of Power Law Models

• Bias is substantially increased for highly nonlinear models

Lec 15 – page 7
 Disattenuation of Power Law Models
• Attenuation can be observed in a
Monte Carlo experiment, like the
situation shown here

• Consider disattenuation to back

out the true underlying power
given the biased fit power and
known error

• Applicable where error is known,

such as small scale
correspondence measurements

Nonlinear Inference with Multivariate

Multiscale Data
• More commonly, multivariate multiscale measurements, such as iron
grade and Bond mill work index are available
• If the relationship is strong enough, simulate the metallurgical property
at the small scale and use nonlinear regression

Lec 15 – page 8
 Nonlinear Inference with Multivariate
Multiscale Data
• Require realizations of Z at the small scale v with:
– the correct spatial regionalization (variogram)
– the correct bivariate distribution with the known correlation
linear variable
• Two evident approaches:
– Drawing realizations of correlated probabilities using
probability field (p-field) simulation
– Semiparametric Bayesian updating (BU) using a familiar
framework from Barnett and Deutsch’s multiple imputation
method

Stochastic Realization of Z(v)

Lec 15 – page 9
 Effect of Bivariate Information
• Increasing bivariate information decreased the mean squared error of
the inferred nonlinear re-expression (mean squared error of fit Z(V)
compared to true Z(V)) when all other parameters held constant

Establishing the Bivariate Distribution

• Bivariate distribution at scale v is
required

• We know how to downscale

univariate distributions:
– Discrete Gaussian model with
mosaic model
– Indirect lognormal model

• Consider downscaling using product

of ratios approach from
Leuangthong, Ortiz & Deutsch

Lec 15 – page 10
Bivariate Distribution Scaling Algorithm
• Scale a bivariate distribution (or up to ~4 dimensional multivariate)
iteratively

1. Fit the bivariate distribution using KDE, KDN, GMM,…

2. Downscale the marginal (univariate) distributions
3. For P iterations determined by a tolerance, update the downscaled
bivariate distribution using the produce of ratios:

4. Compare the marginal distributions

Downscaled Bivariate Distribution

• Demonstrated using case where
the true small scale distribution is
known from scattered data

• After product of ratios

downscaling, univariates match
exactly

• Bivariate distribution is very close

to true bivariate distribution

• Susceptible to curse of
dimensionality

Lec 15 – page 11
 Geometallurgical Case Study
• Several of the described techniques
are demonstrated using a case study
from a copper-molybdenum porphyry
deposit
• Quantifying mill throughput is critical
for this large tonnage operation
– A model of grinding indices and rock
properties is required in addition to
copper and molybdenum grades
• Data consists of 18.5 km of drilling
– Geology includes 13 lithologies and 8
alterations
– Modeled domain of highly altered, high
grade, and immediately surrounding
host rock is displayed
• Not the focus here

23

Data Inventory
• Data continued:
– Assays: copper, molybdenum, silver, iron, sulfur, acid soluble copper and
cyanide (2 m intervals)
• Interval data also included rock quality designation (RQD), core recovery,
point load test strength and fractures frequencies
– Metallurgical data: communition and whole rock analysis (30m intervals)
• Includes BMWi, semi-autogeneous grinding (SAG) and critical semi-
autogeneous grinding index (Abx)
• Data available on 2m intervals are composited to 15m (bench scale) for
modeling
• Due to the absence of experimental data comparing assay and
geometallurgical variables, the two are modeled separately
– Comparison will be made between using models that are constructed using
30m intervals and downscaled 15m intervals for geometallurgical variables

24

Lec 15 – page 12
 Model Workflow
• Capturing parameter and data
uncertainty at several steps,
the workflow may be
summarized as:
1. Normal score the variables
2. Impute missing values
3. Simulate using ICCK (with
SGSIM) using supersecondary
variables
4. Correct realizations to have
the correct histogram
5. Check reproduction

25

Model Checking
• Model checking typically includes:
– Visual inspection of realization and e-type
– Histogram reproduction
– Multivariate reproduction
– Variogram reproduction

Realization E-type

26

Lec 15 – page 13
 Model Checking (2)

Data

Realization

27

Optimal Pit Uncertainty

• An economic model is created that includes costs for crushing and
grinding as a function of grinding indices and density, and prices for the
final concentrate as a function of both copper and molybdenum grade
– Lerchs-Grossmann uses this model, which penalizes low-throughput ore
corresponding to low Axb values
• Workflow is summarized as:
1. Convert realizations of modeled variables to economic value
2. Execute Lerchs-Grossman on the economic value realizations to generate
20 pit realizations
3. Probability that a block is optimally mined in 50% of the pits is contoured to
provide a simplified measure of the optimal pit

28

Lec 15 – page 14
 Variability of Throughput Rate
• Regions of high Axb corresponding to less energy intensive ore and low
Axb (more energy intensive ore) are observed in the pit limits
• The temporal variation of Axb is examined in an intermediate pit
expansion over a period of 300 operating days
– An extraction sequence is applied to 20 high-resolution realizations
– The average daily Axb was calculated for each day

29

Downscaling
• The realizations of Axb that are used in the previous slide are
constructed using the 30m measurements of Axb, not
measurements on a 15m scale
• Consider downscaling to 15m
– Use an assumption of linearity since data is not available to support
non-linear averaging schema
• Downscaling of the variogram displayed below
– Lead to a variance increase of 30%

30

Lec 15 – page 15
 Increased Variability of Throughput
• The downscaled histogram is
stable and reproduces the
increase in variance of 30%
• The resulting realizations of Axb
show the effect of the
substantially increased variance

30m Scale 50m Scale

31

Review of Main Points

• Nonlinearity is important – high degree of mixing

during comminution
• Prefer inference with direct experimentation
• Multiscale measurements may be used, and if
error is known disattenuated
• Multiscale correlated bivariate measurements
may be used if the relationship contains
sufficient information
• If required downscale the bivariate distribution
using product of ratios approach

Lec 15 – page 16
Geostatistics for Mineral Resources Uncertainty Assessment

Post Processing and Data Spacing

• Post processing
• Principles of data spacing

Reporting
• Practitioners may wonder how to report in presence of multiple
realizations – people still want just one model!?!
• Reporting within one zone (time period, bench, stope, …):

• In presence of multiple realizations, we could report common

quantiles (P10/P50/P90) or (P5/P50/P95):

• These tables are for one zone. The zones could be changed to achieve
certain targets for tonnage, strip ratio, metal content…
• Assigning probability intervals to all tonnage/grade numbers is an important
step

Lec 16 – page 1
 By Time Period
• Present uncertainty in different
time periods
• Uncertainty in NPV/ROI-type
statistics would require
realizations, which we have
• Schematic to the right shows
error bounds and multiple
realizations
• Should exceed P90 10% of the
time and fall below P10 10% of the
time
• It is fairly standard to assemble
uncertainty in this manner
– What about the source of
uncertainty? RT or grade?
– Sensitivity versus Uncertainty

Uncertainty
• P10/P50/P90 results for the
10 zones
• Realizations are shown
• Results are consistent:
– 10% above P90
– 10% below P10
• Could process through
cash flow analysis…

• Summarize by expected
value and show
uncertainty

Lec 16 – page 2
 Review of the Approach
The Efficient Frontier
• Originally proposed by H. Markowitz (1952) as a way to select between
investment portfolios

The Efficient Frontier and pit optimization

• Rank options based on expected value vs risk
• Maximum expected value option
– Provides maximum acceptable risk
• For a specific level of risk
– Best choice maximizes expected value

* Figure is a simplified example of the efficient frontier originally presented by H. Markowitz in 1952 in “Portfolio Selection"

A Case Study with Traditional Results

Models used in Case Study*
• Stochastic Models
– Based off of 75,980 samples, 3 variables, and 7 domains defined from 18
rock types
– Expected Block Values range from -2 to 14 (1% outliers go higher)
– Grid size NX = 144, NY=110, NZ=21 (Total of 332,640 block)

• Locally optimal results

– Can find a maximum E{value}
– Can find a frontier of pit options
* Stochastic grade models prepared by Felipe A. Pinto for a Master of Science Thesis (2016) using data provided by Teck Resources Limited.

Lec 16 – page 3
 The Efficient Frontier
A suggested approach for using the efficient frontier methodology
1. Find the pits along the efficient frontier
2. Review and understand the changes between the pits along the frontier
3. Use the efficient frontier and the review of changes to better inform
future decision making processes.

Pits Along the Efficient Frontier

Efficient Frontier Plot
1. Finding the pits along the efficient
frontier

Logic Check the Results:

• Is the maximum expected value pit
similar to the pit optimized on the
expected block value model?
• Do the pits optimized globally (over all
models) out perform pits optimized
locally (over individual models)?

Lec 16 – page 4
 Surfaces Along the Frontier
Efficient Frontier Plot
2. Review the changes in the pits
along the efficient frontier

Reviewing surface plots give some insight

However what is the story here?

Depth of Pit
Depth of Pit
What is Changing? (Cross-Sections)
How much do the pits change?
Surface Plots
Difference in Depth

• Dashed lines represent

cross-sections
Cross-Sections
Difference in Depth

• Pit Outlines
• Probability to be Ore
Decimal Probability
Decimal Probability

South North

Lec 16 – page 5
 Small 2D Examples
Small 2D examples show some similarities and hint at other shapes to the
frontier
Similar Features:
• Smooth slope where sides of the pit
are being shaved away

Other Features:
• Strong Jumps in the Frontier
where larger portions
of the pit are dropped

Principles of Data Spacing

and Uncertainty
• Geologic data is collected to reduce uncertainty, leading to improved
technical decisions

• Uncertainty will always exist to some extent

– Too much uncertainty leads to poor decisions and reduced project value
– Too much data leads to acquisition costs that are not recovered via
improved decisions

• Data spacing and uncertainty studies performed to assess the ‘best’ or

target data spacing. Addresses common questions that include:
1. How should we quantify uncertainty?
2. How much does uncertainty reduce with more data?
3. What factors control uncertainty other than data spacing?
4. How many do we need to know how many we will need?
5. What uncertainty should we target?

Lec 16 – page 6
 Quantifying Uncertainty
• Principle: uncertainty can be quantified with geostatistical simulation and
is scale dependent.

• Practitioners must work to incorporate all known sources of uncertainty

– Parameter uncertainty, measurement uncertainty, etc.
– Minimize unknown unknowns

Sources of Uncertainty
• Principle: uncertainty
can be explained by
data spacing and other
local factors.

Well Spacing NN Distance Conditional Mean

Lec 16 – page 7
Predicting How Many Data are Needed
• Principle: local factors can be used for
estimating the data spacing that yields a
target uncertainty.

How Many Do We Need, to Know How

Many We Will Need?
• Principle: there is
a minimum
number of data
spacing that is
required to
understand the
local factors.

• Studied with
resampling and
resimulation

Lec 16 – page 8
 Optimal Spacing
• Principle: the cost of uncertainty should be estimated to determine a
target data spacing and associated uncertainty.

Reasonable Spacing
• Principle: if the cost of uncertainty cannot be estimated, characteristics of
data spacing and uncertainty can be used to determine a reasonable
target.

Lec 16 – page 9
Geostatistics for Mineral Resources Uncertainty Assessment

Resource Classification

• Purpose
• Geometric criteria
• Probabilistic support
• Recommendations

Background
• Canadian Securities Commissions (ASC, BCSC, CVMQ, OSC) and the
Toronto Stock Exchange (TSE) have adopted National Instrument 43-
101: Standards of Disclosure for Mineral Projects.
• NI 43-101 applies to all oral statements and written disclosure of
scientific or technical information, including disclosure of a mineral
resource or mineral reserve.
• NI 43-101 defers to the Canadian Institute of Mining, Metallurgy and
Petroleum for definitions and guidelines. In particular, the CIM
Standards on Mineral Resources and Reserves Definitions and
Guidelines adopted by CIM Council on August 20, 2000.
• Council of Mining and Metallurgical Institutes (CMMI) of which CIM is a
member, have developed a Resource/Reserve classification, definition
and reporting system that is similar for Australia, Canada, Great Britain,
South Africa and the United States.
• The Joint Ore Reserves Committee (JORC) of the Australasian Institute
of Mining and Metallurgy (AusIMM) and the Australian Institute of
Geoscientists and the Minerals Council of Australia has received broad
international acceptance.

Lec 17 – page 1
 Reserve Definitions
• A Mineral Reserve is the economically mineable part of a Measured or Indicated
Mineral Resource demonstrated by at least a Preliminary Feasibility Study. This
Study must include adequate information on mining, processing, metallurgical,
economic and other relevant factors that demonstrate, at the time of reporting,
that economic extraction can be justified. A Mineral Reserve includes diluting
materials and allowances for losses that may occur when the material is mined.
– A ‘Proven Mineral Reserve’ is the economically mineable part of a Measured Mineral
Resource demonstrated by at least a Preliminary Feasibility Study. This Study must
include adequate information on mining, processing, metallurgical, economic, and
other relevant factors that demonstrate, at the time of reporting, that economic
extraction is justified.
– A ‘Probable Mineral Reserve’ is the economically mineable part of an Indicated, and in
some circumstances a Measured Mineral Resource demonstrated by at least a
Preliminary Feasibility Study. This Study must include adequate information on mining,
processing, metallurgical, economic, and other relevant factors that demonstrate, at
the time of reporting, that economic extraction can be justified.

Historical Approaches

• Expert assessment of geological continuity

• Distance from a sample(s)
• Sample density in the vicinity of a block
• Geometric configuration of data available to estimate a block
• Kriging variance or relative kriging variance

Lec 17 – page 2
 Geometric Measures

• Geometric measures such as drillhole spacing, drillhole density

and closeness to the nearest drillhole are direct measures of the
amount of data available

and

10000 L12  L22  (2r ) 2

d  L1  L2  10000
L1  L2  
 2  d
L12  L22
r
4

• They are related

Calculating Geometric Measures

• Calculate the density of data per unit volume
• Choice of thresholds depend on:
– Industry-standard practice in the country and geologic province
– Experience from similar deposit types
– Calibration with uncertainty quantified by geostatistical calculations
– Expert judgement of the competent or qualified person

10000  n 10000
d L
  rs2 d

Lec 17 – page 3
 Uncertainty
• Geometric methods for classification are understandable, but do not
give an actual measure of uncertainty or risk.
• Professionals in ore reserve estimation increasingly want to quantify
the uncertainty/risk in their estimates.
• The JORC code:

Application to Mining
• Volume:
– Large variability at small scale
– Variability reduces at large scale
– Volume translates to “time”

• Geological continuity:
– Greater continuity – less variability
– Continuity depends on facies/rock type

• Data configuration
– Drilling is not uniform
– Areas of greater drillhole density will
have less uncertainty

• High/low grades
– Variability/uncertainty often depends
on grade level
– High grades can be more variable

Lec 17 – page 4
 Classification of Resources
and Reserves
• Three aspects of probabilistic classification:
1. Volume
2. Measure of “+/-” uncertainty
3. Probability to be within the “+/-” measure of uncertainty
• Format for uncertainty reporting is clear and understandable, e.g.:
– monthly production volumes where the true grade will be within 15% of the
predicted grade 90% of the time are defined as measured
– quarterly production volumes where the true grade will be within 15% of
the predicted grade 90% of the time are defined as indicated
• Drillhole density and other geometric measures are understandable,
but do not really tell us the uncertainty or risk associated with a
prediction
• There are no established rules or guidelines to decide on these three
parameters; that remains in the hands of the Qualified Person

1. Volume
• “Volumes relevant to technical and economic evaluation”
• Consider a regular grid of volumes; classify central point by specified
volume around it rather than complex volume within a mine plan that
could change

• Production and economic indicators are particularly relevant for

monthly and quarterly time periods
• Volume for measured/proven could be smaller than the volume for
indicated/probable
• Consider block volumes or tonnages that are consistent with
reasonable time periods, e.g.:
– Monthly for measured/proven
– Quarterly for indicated/probable

Lec 17 – page 5
 2. Measure of “+/-” uncertainty
• Analytical procedures can achieve a very tight tolerance:
– 103 grams at most
– the entire sample is accessible for measurement
– independent repeat measurements reduce uncertainty 2=2s/n
• Mineral deposits are different:
– 1012 grams or more
– less than 1 billionth of the sample available for analysis
– data are correlated because of the geological setting
– must predict relatively small local volumes between available data
• Practical guidelines for +/- uncertainty:
– 5 % lower limit – very continuous orebody
– 25% upper limit – prediction quite poor indeed with more than 25% error
– 15% is a suitable compromise and is becoming accepted

3. Probability to be within the“+/-”

measure of uncertainty
• Consider the same factors as +/- measure of uncertainty
– 80%: P10 to P90
– 90%: P05 to P95
• Must meet some minimum threshold for classification (the uncertainty
on the right is too great):

Probability 95% 80% 50%

to be within:

Lec 17 – page 6
 Schematic Illustration

1. Volume relevant for

technical and economic
evaluation

2. Predicted grade: z*
and chosen measure of
-15% z* +15%
uncertainty: +/- 15%

3. Probability to be within
measure of uncertainty

Does probability meet

criteria?

Uncertainty
• Predictions (z* values) can be from geostatistical or more traditional methods
• Geostatistical procedures are used to construct probability distributions of
uncertainty:
– parameters vary locally and within facies
– there are a number of different techniques including simulation
• Uncertainty is predicted at different scales (small blocks, approx. monthly
volumes, approx. quarterly volumes,…)
• Distributions of uncertainty can be checked:
– Predict uncertainty at locations where we have drillholes
– Construct different probability intervals (e.g., 10, 50, and 90%)
– Count the number of times the true values is in these intervals
– Should be in the interval the correct percentage (10, 50, and 90%)
• Geostatistical procedures have been shown to reliably predict the uncertainty
due to incomplete sampling

Lec 17 – page 7
 Concerns with Probabilistic Approach
• There are four main concerns:
1. Uncertainty is model-dependent and stationarity-dependent.
Uncertainty can be changed dramatically be minor changes to these
decisions.
2. Many parameters affect the distributions of uncertainty in a non-
intuitive and non-transparent manner.
3. Uncertainty in the histogram and
spatial continuity parameters is not
commonly considered, but can have
a large affect on large mining-scale
uncertainty.
4. Choosing the parameters of uncertainty
for classification is cannot be universal
and is very deposit specific
• These concerns are serious enough that a geometric-based
classification scheme that is backed up by uncertainty may be the
most robust

Example of Probabilistic Classification

• Note the limited

number of drillholes
• A reference histogram
and variogram are
used
• The probability to be
within +/-50% of the
estimated grade could
be used for
classification

Lec 17 – page 8
 More Drilling and a Change in Modeling

• More drillholes (190 vs

21 before)
• Different decision of
stationarity
• New calculation of
probabilities
• The probability to be in
interval has mostly
increased, but 16% of
the probabilities are
lower!

Example of the Importance of Parameters

• Real dataset with

variograms fitted
with different
nugget effects
• Let’s see
realizations and
probabilistic criteria

Lec 17 – page 9
 More on Example of the Importance of
Parameters
• Realizations and
probability to be
within 50% of
expected
• Significant change
in probabilities (see
below)

Discussion
• Geometric criteria are easy to understand and are recommended in most
cases
• Three aspects of probabilistic classification:
1. Volume (monthly and quarterly)
2. Measure of “+/-” uncertainty (15%)
3. Probability to be within the “+/-” measure of uncertainty (80%)
• Quantitative uncertainty should be used to support choice of geometric
criteria
• There will always be a need for Qualified Persons to render expert
opinions in different areas related to geological interpretation, sampling,
and estimation procedures.
• The rather loose definitions in current usage will be tightened up by
lending probabilistic meaning to classificaiton that is easily understood
by technical staff, management, and the investment community.

Lec 17 – page 10
Geostatistics for Mineral Resources Uncertainty Assessment

Final Remarks and Case Studies

• Remarks
• Olympic Dam

Olympic Dam Case Study

• Conventional resource estimation focuses on one or a few metals
– This discussion using a mining context, though similar themes apply to
petroleum
• Increasingly, it is becoming important to understand other characteristics
of the ore that affect processing performance and recovery
– Problem becomes potentially massively multivariate
• Consider BHP Billiton’s Olympic Dam project in South Australia
– Total of 204 measured variables that may or may not relate to 6 plant
performance variables

Lec 18 – page 1
 Prediction Model Steps
• Following Boisvert et al. (2013), the first section of the Olympic Dam
project requires predicting plant performance from measured variables:
1. Remove unimportant and redundant variables.
2. Quantile to quantile univariate transformation to a Gaussian distribution.
3. Merge the variables (level 1). This step reduces the 112 input variables to
23 merged variables.
4. Merge the variables (level 2). This step reduces the 23 merged variables to
4.
5. Regression on the 4 variables and prediction of the plant performance
variables: DWi, BMWi, Cu recovery, U3O8 recovery, acid consumption, and
net recovery
6. Back transform the estimated variables.
7. Determine uncertainty in the model.
• Geostatistical simulation is performed to spatially model inputs to the
plant prediction model

Boisvert, J.B., Rossi, M.E., Ehrig, K., Deutsch, C.V. (2013) Geometallurgical
Modeling at Olympic Dam Mine, South Australia, Math Geosci, 25: 901-925 3

Amalgamation
• Given the 804 observations, using all 204 variables in a regression
model would lead to overfitting
– Redundant and unimportant variables are identified, reducing the number to
112
– Amalgamation steps are used to further condense the variables to 4
subcategories
• A linear model based on 4 amalgamated variables provides a robust
predictive model that is used for predicting recovery and plant
performance

Lec 18 – page 2
 Amalgamation (2)
Merged 1-3: Head Grade Variables
LEVEL 1: Merged 4-6: Mineralogy
Reduce 112 secondary variables to 23 variables
(16 merged variables + 7 variables retained) Merged 7-16: Association Variables
Retained Cu, Ag, Au, S, U3O8, SG, ratios
Retained Merged_1 Merged_2 Merged_3 Merged_4 Merged_5 Merged_6 Merged_7 Merged_8 Merged_9
Cu(wt%) Co(ppm) Ba(wt%) La(wt%) Uran_Wt% Chal_Wt% Sul_Wt% Bran_Pyr_assoc Cof_Bran_assoc Uran_Cof_assoc
U3O8(ppm) Mo(ppm) Fe(wt%) Mg(wt%) Cof_Wt% Born_Wt% A_Sol_Wt% Bran_Chalcopy_assoc Cof_Uran_assoc Uran_Chalcopy_assoc
SG Pb(ppm) Al(wt%) Mn(wt%) Bran_Wt% Chal_Wt% A_Insol_Wt% Bran_Bornite_assoc Cof_Pyr_assoc Uran_Bornite_assoc
K:Al Zn(ppm) Si(wt%) Na(wt%) Pyr_Wt% Bran_Chalcocite_assoc Cof_Chalcopy_assoc Uran_A_Sol_assoc
Ag(ppm) K(wt%) P(wt%) Bran_A_Sol_assoc Cof_Chalcocite_assoc Uran_A_Insol_assoc
Au(ppm) Ca(wt%) Ti(wt%) Bran_A_Insol_assoc Cof_Sulphides_assoc
Badj%S S(wt%) Ce(wt%) Bran_Free_Surf_assoc Cof_A_Sol_assoc
CO2(wt%) Cof_A_Insol_assoc
F(wt%) Cof_Free_Surf_assoc
Merged_10 Merged_11 Merged_12 Merged_13 Merged_14 Merged_15 Merged_16
Pyr_Cof_assoc Chalcopy_Bran_assoc Bornite_Cof_assoc Chalcocite_Chalcopy_assoc Sulphides_Uran_assoc A_Sol_Bran_assoc A_Insol_Bran_assoc
Pyr_Uran_assoc Chalcopy_Cof_assoc Bornite_Pyr_assoc Chalcocite_Bornite_assoc Sulphides_Pyr_assoc A_Sol_Cof_assoc A_Insol_Cof_assoc
Pyr_Chalcopy_assoc Chalcopy_Uran_assoc Bornite_Chalcopy_assoc Chalcocite_Sulphides_assoc Sulphides_Chalcopy_assoc A_Sol_Uran_assoc A_Insol_Uran_assoc
Pyr_Sulphides_assoc Chalcopy_Pyr_assoc Bornite_Chalcocite_assoc Chalcocite_A_Sol_assoc Sulphides_Bornite_assoc A_Sol_Pyr_assoc A_Insol_Chalcopy_assoc
Pyr_A_Sol_assoc Chalcopy_Bornite_assoc Bornite_Sulphides_assoc Chalcocite_A_Insol_assoc Sulphides_A_Sol_assoc A_Sol_Chalcopy_assoc A_Insol_Bornite_assoc
Pyr_Free_Surf_assoc Chalcopy_Chalcocite_assoc Bornite_A_Sol_assoc Chalcocite_Free_Surf_assoc Sulphides_A_Insol_assoc A_Sol_Bornite_assoc A_Insol_Sulphides_assoc
Chalcopy_Sulphides_assoc Bornite_A_Insol_assoc A_Sol_Chalcocite_assoc A_Insol_A_Sol_assoc
Chalcopy_A_Sol_assoc Bornite_Free_Surf_assoc A_Sol_Sulphides_assoc A_Insol_Free_Surf_assoc
Chalcopy_A_Insol_assoc A_Sol_A_Insol_assoc
Chalcopy_Free_Surf_assoc A_Sol_Free_Surf_assoc

A B C D
Cu(wt%) Merged_1 Merged_4 Merged_7
U3O8(ppm) Merged_2 Merged_5 Merged_8 How to merge variables:
LEVEL 2: SG Merged_3 Merged_6 Merged_9
4 final variables K:Al Merged_10
Ag(ppm) Merged_11
Au(ppm) Merged_12
Badj%S Merged_13
Merged_14
Merged_15
Merged_16

Variable A contains individual variables retained.

Variable B contains the remainder of the head assays.
Variable C contains all mineralogy variables.
Variable D contains all association variables.

Response Surface
• Fit a response surface to six performance (metallurgical) variables
Variable A contains individual variables retained.
Variable B contains the remainder of the head assays.
Variable C contains all mineralogy variables.
Variable D contains all association variables.

• The variables are merged differently for each response variable

• Check by cross validation
• Important to understand which of the 112 input variables are actually
being used to achieve the fit

Lec 18 – page 3
 Sensitivity Analysis

White – Head Assays

Gray – Associations
Red – Mineralogy
Black – SG
7

Input Variables Output Variable

1st Variable 2nd Variable 3rd Variable Confidence Count
S + Ti + Na = DWi 0.90 103
A_Insol_Chalcopy + Chalcopy_Wt% + Na = DWi 0.91 100
S + Chalcopy_Wt% + Na = DWi 0.93 98
A_Insol_Free + S + Na = DWi 0.93 96

Rule Induction
SG + Chalcopy_Wt% + Na = DWi 0.93 96
Na + Fe = DWi 0.91 110
Na + Fe + Bran_Pyr = DWi 0.91 110
Na + Fe + Sulph_Born = DWi 0.91 110
Na + Fe + Chalci_Free = DWi 0.91 110
Na + Fe + Sulph_Uran = DWi 0.91 110
Ti + A_Sol_Chalcopy + A_Sol_A_Insol = BMWi 0.90 73
Si + Chalcopy_Free + Uran_Wt% = BMWi 0.91 66
Si + Chalcopy_Free + A_Sol_Uran = BMWi 0.92 66
Chalcopy_Free + A_Sol_Uran + SG = BMWi 0.92 66
Ti + Born_A_Sol + Mn = BMWi 0.91 65

• A rule is an association of 2 or
P + Cof_Uran + Ti = BMWi 0.91 75
Cof_Free + Sulph_A_Insol + Ti = BMWi 0.90 71
P + Chalcopy_Pyr + Ti = BMWi 0.91 63

3 input variables that predict a P

P
Pyr_Wt%
+
+
+
A_Insol_Free
Uran_Wt%
Chalcopy_Free
+
+
+
Uran_Wt%
A_Insol_Wt%
Ti
=
=
=
BMWi
BMWi
CuRec
0.91
0.91
0.92
58
57
51

response with: A_Sol_Free

A_Insol_Sulph
+
+
A_Sol_Sulph
Pyr_Sulph
+ A_Sol_Chalcopy =
+ Ti =
CuRec
CuRec
0.90
0.92
50
47
A_Sol_Sulph + Chalcopy_Bran + Chalcopy_Wt% = CuRec 0.91 43

– High confidence (probability) U3O8

Chalcopy_Wt%
+
+
Pyr_Wt%
Bran_Wt%
+ Chalcopy_Free =
+ A_Sol_Born =
CuRec
CuRec
0.93
0.94
43
31
Ti + Chalcopy_Wt% + Sulph_Pyr = CuRec 0.93 30

– High count (coverage) Chalcopy_Wt%

Chalcopy_Wt%
+
+
Bran_Wt%
Bran_Wt%
+ Born_Wt%
+ Chalcopy_Born =
= CuRec
CuRec
0.97
0.97
30
30
Chalcopy_Sulph + Born_Chalcopy + Cof_Wt% = CuRec 0.90 30

• Figure to right shows 60 rules: Cu

Chalci_Wt%
Cu
+
+
+
A_Sol_Chalcopy
Chalcopy_Wt%
Uran_Cof
+ Chalcopy_A_Sol =
+
+
Mg
Fe
=
=
U3O8Rec
U3O8Rec
U3O8Rec
0.91
0.90
0.90
42
41
41

– Six response variables

Cu + A_Insol_Wt% + Mg = U3O8Rec 0.90 40
Cu + Uran_Cof + Na = U3O8Rec 0.92 39
Na + Ti + Cu = U3O8Rec 0.91 66

– Five positive and five negative

S + A_Sol_Born + Cu = U3O8Rec 0.92 66
Na + A_Insol_Uran + Cu = U3O8Rec 0.92 65
S + Born_Wt% + Cu = U3O8Rec 0.91 65

– Red is high (positive)

S + Chalcopy_Born + Cu = U3O8Rec 0.91 64
A_Insol_A_Sol + Pyr_Wt% + Sulph_Born = Acid_consumption 0.94 132
A_Insol_A_Sol + Pyr_Wt% + Bran_Pyr = Acid_consumption 0.94 132

– Black is low (negative)

A_Insol_A_Sol + Pyr_Wt% + Sulph_Uran = Acid_consumption 0.94 132
A_Insol_A_Sol + Pyr_Wt% + Chalci_Free = Acid_consumption 0.94 132
A_Insol_A_Sol + Pyr_Wt% = Acid_consumption 0.94 132
A_Sol_Wt% + Chalcopy_A_Sol + Sulph_Born = Acid_consumption 0.90 140
• Results reviewed with A_Sol_Wt%
A_Sol_Wt%
+
+
Chalcopy_A_Sol
Chalcopy_A_Sol
+
+
Chalci_Free
Bran_Pyr
=
=
Acid_consumption
Acid_consumption
0.90
0.90
140
140

geologists and metallurgists

A_Sol_Wt% + Chalcopy_A_Sol + Sulph_Uran = Acid_consumption 0.90 140
A_Sol_Wt% + Chalcopy_A_Sol = Acid_consumption 0.90 140
A_Sol_Born + A_Insol_Uran + Bran_Pyr = %extr-mass_loss 0.93 106

with some interesting insight A_Sol_Born

A_Sol_Born
A_Sol_Born
+
+
+
A_Insol_Uran
A_Insol_Uran
A_Insol_Uran
+
+
+
Sulph_Born
Chalci_Free
Sulph_Uran
=
=
=
%extr-mass_loss
%extr-mass_loss
%extr-mass_loss
0.93
0.93
0.93
106
106
106
A_Sol_Born + A_Insol_Uran = %extr-mass_loss 0.93 106
Cof_Uran + Chalcopy_Born + A_Sol_Born = %extr-mass_loss 0.91 85
A_Insol_Uran + Bran_A_Sol + Cof_Uran = %extr-mass_loss 0.92 84
A_Insol_Born
A_Insol_Uran
+
+
Cof_Uran
A_Insol_Born
+ Chalcopy_Born =
+ Cof_Uran =
%extr-mass_loss
%extr-mass_loss
8
0.91
0.90
84
83
F + A_Insol_Born + A_Sol_Born = %extr-mass_loss 0.90 82

Lec 18 – page 4
 Spatially Modeling of Variables
• 112 variables must be spatially mapped to utilize the developed models
• There are two main difficulties:
– the compositional nature of the variables must be accounted for
– many of the variables are correlated and require methodologies that can be
applied to a large number of variables but are effective in reproducing the
multivariate relationships.
• 27M cell model creates computational challenges for 112 variables

High Level Workflow

• Variables are separated into three groups:
1. Head grade assays
2. Grains size measurements
3. Mineral associations
• Far more samples exist for the head grade variables → modeled first
• Grain size and association variables are modeled using head grade
realizations as secondary information

10

Lec 18 – page 5
 Head Grade Simulation
• Head grades are simulated using logratios and PCA
– Updated workflow would likely use PPMT and/or MAF

Data Realization

11

Grain Size Simulation

• Merge head grade assays
into variable that predicts
grain size variables
• Simulate grain sizes
conditional to simulated
head grades and other
grain sizes Realization

Data

12

Lec 18 – page 6
Spatial Prediction of Plant Performance
• Maps of Copper (left) and Uranium (right) recovery

13

Lec 18 – page 7