EVALUATION OF METALLURGICAL RECOVERY

FACTORS FOR DIAMONDS RECOVERED FROM

KIMBERLITES

Stephen John Coward

School of Civil, Environmental and Mining Engineering

The University of Adelaide

This dissertation is submitted for the degree of Doctor of Philosophy

April 2020
 Evaluation of Metallurgical Recovery Factors

ABSTRACT

Extraction and recovery of diamonds requires that the host rock, kimberlite, is
fragmented to liberate and recover the contained diamonds. Optimal recovery
requires trade-offs to be made between maximising liberation, minimising diamond
breakage or loss and cost of recovery.

Effective fragmentation and recovery are not only dependent on the comminution
and recovery techniques used but are also a function of the interactions between the
diamond characteristics, the host rock properties and the technology used to crush
the kimberlite and recover the diamonds. Prior approaches have been limited by a
disregard for these relationships and how they change in response to variable
kimberlite and diamond characteristics and their impact on diamond recovery.
Incorrect recovery estimation impacts negatively on the evaluation, design and
operation of diamond mining projects.

This research develops and demonstrates methods to collect and spatially estimate
relevant orebody characteristics that impact on diamond liberation and subsequent
recovery. These characteristics are used in an integrated value chain model to
quantify the variability and uncertainty of diamond recovery. The use of this
technique is demonstrated in two case studies.

The benefits of this approach include improved evaluation of diamond projects,

development of better design and operational strategies and will improve not only
the resilience of individual diamond projects, but also the performance and
economics of the diamond mining industry.

Key words – diamond, diamond size distribution, diamond recovery, kimberlite,

Value chain model, process simulation, Liberation, comminution
 STATEMENT OF ORIGINALITY

Submitted by Stephen Coward to the University of Adelaide as a thesis for the degree
of Doctor of Philosophy December 2019.

I certify that this work contains no material which has been accepted for the award
of any other degree or diploma in my name, in any university or other tertiary
institution and, to the best of my knowledge and belief, contains no material
previously published or written by another person, except where due reference has
been made in the text. In addition, I certify that no part of this work will, in the future,
be used in a submission in my name, for any other degree or diploma in any
university or other tertiary institution without the prior approval of the University
of Adelaide and where applicable, any partner institution responsible for the joint-
award of this degree.

I give permission for the digital version of my thesis to be made available on the web,
via the University’s digital research repository, the Library Search and also through
web search engines, unless permission has been granted by the University to restrict
access for a period of time.

I acknowledge the support I have received for my research through the provision of
an Australian Government Research Training Program Scholarship.

2019-12-12

Stephen Coward Date

 ACKNOWLEDGEMENTS

I am sincerely grateful to my family, friends and colleagues who have made this
research possible and have helped in so many ways to develop and produce this
thesis.

This research has taken place over a substantial period and has involved many
collaborators who have helped to shape both the thinking behind the research and
inform various avenues to resolve the research questions.

There are several groups of people who were particularly influential in the various
phases of this research including the following:

Research Supervision at the University of Leeds and the University of

Adelaide:

Professor Peter Alan Dowd who as principal supervisor has provided

incredible patience and guidance throughout the process.

Minrad Extended Team;

Wynand Kleingeld, who was the initiator of the Wells research unit, and other
members of the team including Niall Young, Johan Ferreira, Rob Pearce, Chris Prins,
Ina Dohm, and Tessa Dodkin.

Matthew Field, who's collaboration, understanding and insights in the field of

kimberlite geology provided a solid foundation for the design and execution of the
sampling campaign that provided data for this research.

Grant Nicholas who played a material role in the formulation of the initial scope of
the integrated model .

Special mention must be made of Chris Prins contribution and assistance with the
finalisation of the thesis.

QG Team:

John Vann and Scott Jackson.

École nationale supérieure des mines de Paris:

Allan Galli, Margaret Armstrong, and Christian Lantuéjoul.

Academics:

Roussos Dimitrakopoulos, Kim Esbensen, David Vose.

Two anonymous reviewers who helped improve the content substantially.

Acknowledgement is also given to De Beers South Africa, Debswana, De Beers

Marine (Pty) Ltd, De Beers Marine: Namibia, De Beers Canada Inc (Exploration) and
Namdeb Diamond Corporation (Pty) Limited for support and making data and
images available.

I would like to express my gratitude to several sponsoring companies including De

Beers Family of Companies, , “Limn” sponsored by Dave Wiseman, Isatis sponsored
by Geovariance, IoGas sponsored by Imdex.

Quantitative Geosciences and Interlaced for the opportunity, generous time,

financial assistance and intellectual stimulation without which this research would
not have been completed.
 GLOSSARY

Term Used Description or Definition

ASTM American Standards for Testing and Materials.

Brownfield Deposits where there is already a history of
mining activity.

BTS Brazilian Tensile strength.

Cpht Carats per hundred tonnes.
HMA Heavy Mineral Analysis.
HPGR High Pressure Grinding Roll, A type of
comminution device that is used to reduce the
size of kimberlite rocks.
Intrinsic Hypothesis Weak stationarity of the increments of a
random function (Olea,1991).
Isotopic data Refers to the availability of spatial sampling
data where the same suite of variables has
been measured at all sampled locations.
kWh/t Kilowatt hours per ton.
LDD Large Diameter Drilling.
Macro diamonds (MIDA) A reference to diamonds that are above 0.5
mm in diameter.
Micro diamonds (MICRO) Refers to diamonds that are less than 0.5mm in
diameter.
NI-43-101 A Canadian Securities Exchange code for the
reporting of mineral resources and reserves.
Ordinary Kriging A linear estimation algorithm for a
regionalised variable that satisfies the intrinsic
hypothesis (Olea,1991).
QA/QC Quality Assurance/Quality Control.
ROC Required Operating Capability.
Material of the earth that may or may not be
mineralised that is incorporated in block
Rock
models and encountered whilst mining.
REP Resource Extension Project.
TKB Tuffasitic Kimberlite Breccia.
SEDAR System for Electronic Document Analysis and
Retrieval, the electronic filing system for the
disclosure documents of scrip issuers across
Canada.
Term Used Description or Definition

Stationarity First order stationarity assumes the mean of a

variable is constant between samples
independent of location;
Second order stationarity is the assumption
that the covariance between two points is the
same for a given distance and direction
regardless of the two points that are chosen.
Test Response variable a variable that is a function of both the rock
and a test that is performed on a rock - e.g. rate
of dissolution.
Test Variables parameters and characteristics of a test being
performed that have an impact on the outcome
of a rock test - e.g. the rate of loading during
compressive strength tests.
t10 Percentage passing one-tenth of a particle’s
original size for a given energy input.

UCS Uniaxial compressive strength.

Unit Process response variable A variable that represents the performance of
a full-scale mineral process, e.g. the size
distribution of rock produced by a crusher.

Waste Rock material that may or may not contain

mineral of interest, but that is deemed to have
insufficient value to be selected for treatment.
Young’s Modulus Within the limits of elasticity, the ratio of the
linear stress to the linear strain is termed the
modulus of elasticity or Young's Modulus and
may be written as E =(Stress/Strain). It is this
property that determines how much a
specimen will deform under load and is used
to estimate the strength of rock.
 Contents

1 INTRODUCTION ................................................................................................................ 22
1.1 CURRENT APPROACHES USED TO EVALUATE DIAMOND MINING PROJECTS ............................ 23
1.2 SPECIFIC CHALLENGES FACED IN EVALUATING DIAMOND-MINING PROJECTS ....................... 25
1.3 DIAMOND PROJECT EVALUATION PRACTICE ............................................................................. 31
1.4 USE OF THE METALLURGICAL RECOVERY FACTOR IN PROJECT EVALUATION ......................... 33
1.5 IMPACT OF RECOVERY PROCESS DESIGN AND OPERATION ON METALLURGICAL RECOVERY. 33
1.6 IMPACTS OF KIMBERLITE PROPERTIES ON TREATMENT PROCESS EFFICIENCY...................... 35
1.7 IMPACT OF DIAMOND PROPERTIES ON THE RECOVERY FACTOR ............................................. 37
1.8 THE SOURCES OF UNCERTAINTY IN ESTIMATING THE RECOVERY FACTOR ............................. 38
1.9 RESEARCH MODEL ....................................................................................................................... 41
1.10 THESIS STRUCTURE .................................................................................................................. 43

2 LITERATURE REVIEW..................................................................................................... 44
2.1 PHYSICAL CHARACTERISTICS OF KIMBERLITES AND DIAMONDS ............................................. 45
2.2 ROCK PROPERTY SAMPLING AND CHARACTERISTIC MEASUREMENTS .................................... 51
2.3 CHALLENGES OF SPATIALLY ESTIMATING PHYSICAL ROCK CHARACTERISTICS ...................... 57
2.4 MINERAL PROCESS MODELS AND SIMULATION PRACTICES ..................................................... 61
2.5 APPROACHES TO EVALUATING UNCERTAINTY AND VARIABILITY IN DIAMOND RECOVERY .. 67
2.6 SUMMARY OF PRACTICES AND LIMITATIONS IDENTIFIED IN THIS REVIEW ............................ 71

3 PROBLEM STATEMENT .................................................................................................. 73

3.1 INTRODUCTION ............................................................................................................................ 73
3.2 CURRENT LIMITATIONS IN THE DERIVATION AND USES OF THE METALLURGICAL DIAMOND
RECOVERY FACTOR. ............................................................................................................................ 73
3.3 THE PROBLEM STATEMENT AND RESEARCH QUESTION ........................................................... 74

4 OREBODY SAMPLING AND COLLECTION OF ROCK CHARACTERISTIC DATA

................................................................................................................................................... 76
4.1 INTRODUCTION ............................................................................................................................ 76
4.2 EXPERIMENTAL DESIGN .............................................................................................................. 81
4.3 DATA COLLECTION PROCESS ...................................................................................................... 93
4.4 DATA ACQUIRED .......................................................................................................................... 94
4.5 OBSERVATIONS ARISING FROM THE DATA COLLECTION ....................................................... 103
4.6 CONCLUSIONS ARISING FROM DATA COLLECTION ................................................................. 104
5 MODELLING PHYSICAL CHARACTERISTICS OF OREBODIES ........................... 106
5.1 INTRODUCTION .........................................................................................................................106
5.2 ESTIMATION OPTIONS CONSIDERED AND COMPARED ...........................................................109
5.3 SELECTION AND PREPARATION OF PHYSICAL CHARACTERISTIC DATA FOR ESTIMATION ..111
5.4 USING MULTIVARIATE TECHNIQUES IN ESTIMATION .............................................................112
5.5 ESTIMATION AND SIMULATION METHODS ..............................................................................116
5.6 GENERIC APPROACH AND CONSIDERATIONS FOR GENERATION OF SUITABLE SPATIAL
MODELS ............................................................................................................................................. 138

6 PROCESS SIMULATION AND ROCK CHARACTERISTICS .................................... 141

6.1 INTRODUCTION .........................................................................................................................141
6.2 MODELLING OF MINING PROCESSES ........................................................................................146
6.3 MODELLING OF DIAMOND TREATMENT PLANT PROCESS ......................................................148
6.4 UNIT PROCESS MODELS ............................................................................................................153
6.5 ESTABLISHING RELATIONSHIPS BETWEEN COMMINUTION AND DIAMOND RECOVERY ......170
6.6 PROCESS SIMULATION AND SENSITIVITY TESTING .................................................................191
6.7 LIMITATIONS OF PROCESSING MODELS ...................................................................................191
6.8 CONCLUSION ..............................................................................................................................192

7 VALUE CHAIN MODELS - CASE STUDY 1 ................................................................. 193

7.1 INTRODUCTION .........................................................................................................................193
7.2 BENEFITS AND LIMITATIONS OF THE METHODOLOGY ...........................................................201

8 VALUE CHAIN MODEL - CASE STUDY 2 ................................................................... 203

8.1 INTRODUCTION .........................................................................................................................203
8.2 THE VALUE CHAIN COMPONENTS ............................................................................................206
8.3 CONSIDERATIONS FOR DEVELOPMENT AND SIMULATION OF INTEGRATED VALUE CHAIN
MODELS ............................................................................................................................................. 211
8.4 CASE STUDY FINDINGS ..............................................................................................................213
8.5 GENERIC FINDINGS FROM THE IEM RELEVANT TO THE DERIVATION OF THE
METALLURGICAL RECOVERY FACTOR ............................................................................................. 216

9 DISCUSSION ..................................................................................................................... 219

9.1 INTRODUCTION .........................................................................................................................219
9.2 A TAXONOMY FOR VARIABLES USED IN RECOVERY FACTOR DETERMINATION ....................219
9.3 INCREASING THE QUALITY OF SPATIAL ESTIMATES -USING PROXY MEASUREMENTS, SCALE-
UP AND TESTING FOR ADDITIVITY OR NON-LINEARITY ................................................................. 224
9.4 DIAMOND METALLURGICAL RECOVERY FACTORS – EXPLORING UNCERTAINTY AND
VARIABILITY BY EFFECTIVE USE OF VALUE CHAIN MODELS .......................................................... 225
9.5 CONCLUSION ..............................................................................................................................226
10 SUMMARY AND CONCLUSIONS ............................................................................... 227
10.1 INTRODUCTION ...................................................................................................................... 227
10.2 CONCLUDING THEMES........................................................................................................... 228
10.3 SPECIFIC CONTRIBUTIONS OF THIS RESEARCH .................................................................... 231
10.4 THE LIMITS OF THIS RESEARCH ............................................................................................ 232
10.5 RECOMMENDED AREAS FOR ADDITIONAL RESEARCH ......................................................... 233

11 REFERENCES ................................................................................................................. 235

12 APPENDICES .................................................................................................................. 246

APPENDIX 1 -EXPERIMENTAL DESIGN AND DATA FROM OREBODY SAMPLING

................................................................................................................................................. 247

APPENDIX 2 -OREBODY CHARACTERISTIC ESTIMATES ...................................... 251

 Evaluation of metallurgical recovery factors for diamonds recovered from kimberlites

LIST OF TABLES

TABLE 1: SELECTION OF SEVERAL DIAMOND MINES SHOWING THE MINE GRADE RECOVERED IN
CPHT, AVERAGE DIAMOND VALUE IN $/CT AND THE AVERAGE REVENUE IN $/TONNE ( PETRA
DIAMONDS, 2019; ANGLO AMERICAN, 2018; DE BEERS GROUP SERVICES 2018 ) ........... 26

TABLE 2: ANALYSIS OF DIAMONDS SIZED BY SCREENING (ADAPTED AFTER FERREIRA 2013) ..... 27

TABLE 3: A DESCRIPTION OF THE COMPONENTS OF VOLCANICLASTIC KIMBERLITES...................... 50

TABLE 4: RELATIONSHIPS BETWEEN ORE PROPERTY AND THE IMPACT ON MINERAL PROCESSING.
........................................................................................................................................................ 63

TABLE 5: A LIST OF PREFERRED VALUATION METHODOLOGIES USED IN SOUTH AFRICA (SILVA AND
MINNITT, 2005) . ........................................................................................................................ 68

TABLE 6: FORMULAE FOR ESTIMATING ELASTIC MODULI OF SOLIDS USING MEASURED ACOUSTIC
VELOCITIES IN ROCK SPECIMENS. ................................................................................................. 89

TABLE 7: SUMMARY OF THE NUMBER OF EACH TYPE OF SAMPLE COLLECTED. ................................ 94

TABLE 8: SUMMARY STATISTICS OF ANALYSIS CARRIED OUT ON PETROLOGICAL SLAB SAMPLES. . 95

TABLE 9: SUMMARY STATISTICS FOR THE UNIAXIAL COMPRESSIVE STRENGTH (UCS) AND
BRAZILIAN TENSILE STRENGTH (BTS) TEST WORK. ................................................................ 96

TABLE 10: SUMMARY STATISTICS OF DROP WEIGHT RESULTS OBTAINED FROM CORES. ................ 97

TABLE 11: SUMMARY OF DOWNHOLE GEOPHYSICAL READINGS TAKEN FROM EACH HOLE DRILLED
......................................................................................................................................................101

TABLE 12: DESCRIPTIVE STATISTICS OF THE DOWNHOLE FORMATION TESTER RESULTS. ...........102

TABLE 13: SUMMARY STATISTICS FOR DROP WEIGHT TEST DATA SHOWING SAMPLE AND
POLYGONAL STATISTICS. .............................................................................................................117

TABLE 14:TABLE AND GRAPHICS OF THE MODELLED VARIOGRAMS FOR P-WAVE VELOCITY, T10
SPECIMEN DENSITY AND T10_E1. .............................................................................................123

TABLE 15: SUMMARY STATISTICS FOR DATA AND FIRST PASS ESTIMATES. ..................................124

12 Stephen Coward - April 2020

 Chapter 1: Introduction

TABLE 16: SUMMARY STATISTICS COMPARING ESTIMATED AND SIMULATED VALUES FOR P-WAVE
VELOCITY AND DROP WEIGHT RESULTS AT THE 5M BLOCK SCALE. ........................................ 129

TABLE 17: OPTIONS FOR USE OF SPATIAL DATA TO ESTIMATE PROCESS RESPONSE VARIABLES. 134

TABLE 18: SUMMARY STATISTICS FOR THE ESTIMATION AND CALCULATION OF BLOCK SCALE
PROPERTIES FOR PATHWAY 1. .................................................................................................. 135

TABLE 19: SUMMARY STATISTICS FOR THE ESTIMATION AND CALCULATION OF BLOCK SCALE
PROPERTIES FOR PATHWAY 2 ................................................................................................... 135

TABLE 20: SUMMARY STATISTICS FOR THE ESTIMATION AND CALCULATION OF BLOCK SCALE
PROPERTIES FOR PATHWAY 3. .................................................................................................. 136

TABLE 21: SUMMARY STATISTICS FOR THE ESTIMATION AND CALCULATION OF BLOCK SCALE
PROPERTIES FOR PATHWAY 4. .................................................................................................. 136

TABLE 22: SUMMARY STATISTICS FOR THE ESTIMATION AND CALCULATION OF BLOCK SCALE
PROPERTIES FOR PATHWAY 5. .................................................................................................. 136

TABLE 23: A TABULAR DEMONSTRATION OF A WHITTEN SELECTION AND BREAKAGE MODEL. . 163

TABLE 24: CALCULATION OF LOCKED DIAMOND CONTENT............................................................. 175

TABLE 25: CALCULATION OF LOCK UP WITH CONSTRAINT PLACED ON THE MAXIMUM SIZE OF
CONTAINED DIAMOND. ............................................................................................................... 176

TABLE 26: LISTING OF DIAMOND SCREEN SIEVE CLASSES AND ASSOCIATED SIEVE APERTURES,
AVERAGE STONE SIZE PER CLASS AND CRITICAL STONE SIZE PER CLASS................................ 180

TABLE 27: DESCRIPTION OF METHOD USED TO CONVERT THE SAMPLED TAILINGS DISTRIBUTION TO
A TOTAL KIMBERLITIC DISTRIBUTION. ...................................................................................... 185

TABLE 28: FITTING OF THE LOGNORMAL MODEL TO THE RECOVERED SIZE DISTRIBUTION. ....... 186

TABLE 29: PROCESS PARAMETERS REQUIRED FOR THE GRANULOMETRY MODEL. ....................... 187

TABLE 30: CALCULATION OF THE MAXIMUM LOCKED DIAMOND SIZE IN KIMBERLITE PARTICLES IN
EACH SIEVE CLASS. ...................................................................................................................... 187

Stephen Coward - April 2020 13

 Evaluation of metallurgical recovery factors for diamonds recovered from kimberlites

TABLE 31: ALLOCATION OF THE LOCKED POTENTIAL IN EACH SIZE CLASS ACCORDING TO THE
PROBABILITIES DERIVED FROM THE RECOVERED DIAMOND SIZE DISTRIBUTION ..................188

TABLE 32: CALCULATION OF THE LOCKED CARAT POTENTIAL PER DIAMOND SIEVE CLASS. ........189

TABLE 33: CALCULATION OF THE LIBERATED AND LOCKED POTENTIAL REVENUE. .....................190

TABLE 34: SUMMARY OF THE CHARACTERISTICS OF THREE SAMPLING CAMPAIGNS AND THAT USED
TO DEFINE THE 'VIRTUAL OREBODY(V-BOD ) .........................................................................196

TABLE 35: THE DESCRIPTIVE STATISTICS FOR THE V-BOD AND EACH SCENARIO FOR GRADE, DYKE
THICKNESS AND THE GEOMETRICAL VARIABILITY OF THE DYKE SURFACE (V1). ..................198

TABLE 36: SCHEMATIC DEPICTION OF TERMINOLOGY USED TO DESCRIBE ASPECTS OF THE

INTEGRATED EVALUATION MODEL. ..........................................................................................204

TABLE 37: SUMMARY OF MAJOR CAPITAL ITEMS PLANNED FOR EACH PHASE. ..............................208

TABLE 38: AVERAGE EXPECTED DMS YIELDS FOR EACH PHASE. ...................................................208

TABLE 39: LIST OF MODEL SETTINGS USED FOR THE FINANCIAL MODEL. ......................................210

TABLE 40: SUMMARY STATISTICS OF SPATIALLY SIMULATED GRADE, DENSITY AND DMS YIELD
VALUES. ........................................................................................................................................213

TABLE 41: SUMMARY STATISTICS OF $/TONNE DEPLETED BY LOBE BASED ON SIMULATED STONE
VALUES. ........................................................................................................................................215

TABLE 42: A LIST OF GUIDING PRINCIPLES FOR ROCK CHARACTERISTIC SAMPLING. ....................223

14 Stephen Coward - April 2020

 Chapter 1: Introduction

LIST OF FIGURES

FIGURE 1: A SCHEMATIC OVERVIEW OF A MINING VALUE CHAIN , AND A BREAKDOWN OF THE

METALLURGICAL RECOVERY FACTOR. ADAPTED AFTER (APPLEYARD, 2001). ..................... 24

FIGURE 2: PLOT OF DIAMOND STONE COUNTS PER SIZE CLASS. ........................................................ 29

FIGURE 3: VARIATION OF THE AVERAGE STONE SIZE ESTIMATED FROM SMALL SAMPLES. ............ 30

FIGURE 4: TYPICAL RELATIONSHIP BETWEEN DIAMOND SIZE AND CUT DIAMOND RETAIL VALUE
(ADAPTED FROM RAPPAPORT, 2009). ....................................................................................... 30

FIGURE 5: COMPARISON OF THE PROPORTION OF CARATS AND PROPORTION OF REVENUE

REPRESENTED BY DIFFERENT DIAMOND SIZE RANGES, FOR THE DIAMOND PARCEL DESCRIBED
IN TABLE 2. ..................................................................................................................................... 32

FIGURE 6: A SCHEMATIC REPRESENTATION OF THE RESEARCH AREAS COVERED IN THIS THESIS. 41

FIGURE 7: MELT VISCOSITY AS A FUNCTION OF TEMPERATURE AT 1 BAR FOR NATURAL MELTS

SPANNING THE COMPOSITIONAL RANGE RHYOLITE TO KOMATIITE. ALL COMPOSITIONS ARE
VOLATILE FREE. THE TEMPERATURE RANGE IS ILLUSTRATIVE OF TYPICAL ERUPTION
TEMPERATURES FOR EACH COMPOSITION – ADAPTED FROM (SPERRA, 2000) AND (SPARKS
ET AL., 2009). ............................................................................................................................... 47

FIGURE 8: A SCHEMATIC VIEW OF A KIMBERLITE PIPE AFTER HAWTHORNE (1975). ................... 48

FIGURE 9: A SCHEMATIC REPRESENTATION OF THE MATERIALS THAT COULD BE FOUND IN A

‘TYPICAL’ KIMBERLITE PIPE, ADAPTED AFTER (FIELD AND SMITH, 1999). ........................... 49

FIGURE 10: A HIERARCHICAL DEPICTION OF THE RELATIONSHIP OF SOURCES OF ERROR ARISING

FROM SAMPLING, ADAPTED AFTER GY (2004). ........................................................................ 54

FIGURE 11: RESULTS OF A NUMBER OF ROCK TESTS ON A RANGE OF ROCK TYPES AFTER COPUR ET
AL. (2003). ................................................................................................................................... 56

FIGURE 12: SCHEMATIC OF THE PROCESS USED TO CREATE A FRAMEWORK TO QUANTIFY THE
IMPACT OF CHANGES TO SAMPLING , ESTIMATING AND SIMULATING KIMBERLITES ON PROCESS
MODELS. .......................................................................................................................................... 59

FIGURE 13: GEOLOGICAL MAP OF VENETIA K2 (AFTER BROWN 2008) SHOWING THE LOCATION OF
THE GEOMET DRILL HOLES (ORANGE CROSS). ............................................................................ 82

Stephen Coward - April 2020 15

 Evaluation of metallurgical recovery factors for diamonds recovered from kimberlites

FIGURE 14: A VIEW OF THE LAYOUT OF THE CORE HOLES DEPICTING THE LOCATION OF THE
SUBSAMPLES. ................................................................................................................................. 82

FIGURE 15:LISTING OF SUBSAMPLES TAKEN FROM CORES THAT WERE FULLY SAMPLED. .............. 83

FIGURE 16: SCHEMATIC REPRESENTATION OF PREPARATION AND ANALYSIS OF PETROLOGICAL

SAMPLES. (GSC=GEOSCIENCE CENTRE LABORATORY, UCT=UNIVERSITY OF CAPE TOWN,
UOB=UNIVERSITY OF BRISTOL, SGS=TESTING LABORATORY IN JOHANNESBURG,
EDINBURGH =EDINBURGH UNIVERSITY). .................................................................................. 90

FIGURE 17: PREPARATION AND ANALYSIS OF GEOMETALLURGICAL SAMPLES. ............................... 91

FIGURE 18: PREPARATION AND TESTING OF GEOTECHNICAL SAMPLES. .......................................... 92

FIGURE 19: XY PROJECTION OF THE LAYOUT OF THE CORE SAMPLES............................................... 93

FIGURE 20: HISTOGRAM OF THE DENSITY OF SAMPLES SUBJECTED TO DROP WEIGHT TESTING.... 97

FIGURE 21: COMPARISON OF HISTOGRAMS OF DROP WEIGHT RESULTS FOR ALL SAMPLES TESTED AT
THREE DIFFERENT INPUT ENERGIES (TOP FIGURE IS LOWEST ENERGY; BOTTOM PANEL IS
HIGHEST ENERGY). ........................................................................................................................ 98

FIGURE 22: HISTOGRAM AND BASE MAP OF SAMPLE DENSITY FOR THE VKBR DOMAIN. .............. 99

FIGURE 23: A HISTOGRAM AND CROSS SECTION OF THE DENSITIES OF THE SAMPLES TAKEN FROM
THE VK FACIES DOMAIN ............................................................................................................... 99

FIGURE 24: COMPARATIVE HISTOGRAMS OF DROP WEIGHT RESPONSES AT THREE DIFFERENT INPUT
ENERGY LEVELS FOR THE VKBR FACIES (LHS) AND THE VK FACIES (RHS). .....................100

FIGURE 25:HISTOGRAM OF “SHORT SPACED DENSITY” SHOWING HIGH READINGS. ......................101

FIGURE 26: A PLOT SHOWING THE THREE READINGS GENERATED BY THE FORMATION HARDNESS
TOOL FOR HOLE DDH357. ........................................................................................................102

FIGURE 27: HISTOGRAM (LHS) AND SEMI-VARIOGRAM (RHS) OF MM OF DEFLECTION

('DIFFERENCE')MEASURED BY THE DOWNHOLE FORMATION TESTING DEVICE. ..................103

FIGURE 28: SCHEMATIC SHOWING PATHWAYS OF GENERATING SPATIAL MODELS OF ROCK

CHARACTERISTICS. ......................................................................................................................108

16 Stephen Coward - April 2020

 Chapter 1: Introduction

FIGURE 29: A PLOT SHOWING THE IMPACT OF THE SIZE OF THE INTERVAL USED ON THE TOTAL
VARIANCE OF THE BULK MODULUS MEASUREMENT. ............................................................... 112

FIGURE 30: A PLOT SHOWING THE LINEAR MODEL DEVELOPED BETWEEN THE FORMATION
HARDNESS TOOL READINGS AND THE UCS VALUES. ............................................................... 113

FIGURE 31: A BOX AND WHISKER PLOT SHOWING THE SAMPLE VALUE USED TO CALIBRATE THE PLS
MODELS ON THE LEFT OF THE PLOT, AND THE ESTIMATE OF UCS DOWN THE HOLE ON THE
RIGHT. THE HEIGHTS OF THE BARS INDICATE ‘GOODNESS OF FIT’ OF THE PARTIAL LEAST
SQUARES MODEL DERIVED FROM MULTIPLE VERSIONS OF THE MODEL USING DIFFERENT
COMBINATIONS OF SAMPLE AND HOLD OUT DATA. .................................................................. 114

FIGURE 32: A PLOT OF MEASURED UCS VALUES, AND MODELS FOR THE SAMPLES BASED ON THE
DOWNHOLE ACOUSTIC SIGNAL AND CUTTER PENETRATION DEPTH. ..................................... 115

FIGURE 33:PCA MODEL OF UCS DOWNHOLE BASED ON A PRINCIPAL COMPONENT ANALYSIS

(PCA) MODEL USING P- WAVE VELOCITY, S WAVE VELOCITY AND LONG DENSITY, USING ONE
PRINCIPAL COMPONENT (UPPER PANEL) AND TWO PRINCIPAL COMPONENTS (LOWER PANEL).
...................................................................................................................................................... 116

FIGURE 34: A 3D REPRESENTATION OF A POLYGONAL ESTIMATE OF DROP WEIGHT VALUES IN THE

0.7M X0.7M X 0.7M GRID. ......................................................................................................... 118

FIGURE 35: AN EXPERIMENTAL AND FITTED MODEL FOR THE VARIOGRAM FOR THE UCS DATA.
...................................................................................................................................................... 119

FIGURE 36: N-S CROSS SECTION OF AREA ESTIMATED SHOWING HIGH UCS ESTIMATES IN HOTTER
COLOURS AND LOWER VALUES IN COOLER COLOURS (LHS) AND A HISTOGRAM OF THE BLOCK
UCS VALUES ESTIMATED. .......................................................................................................... 119

FIGURE 37: HISTOGRAMS OF THE ORIGINAL T10 VALUES (LEFT) AND THE TRANSFORMED VARIABLE
"MASS LESS 5MM IN G/TONNE" (RIGHT). ................................................................................ 124

FIGURE 38: 3 DIMENSIONAL PROJECTIONS OF THE KRIGING OF DROP WEIGHT VALUES (LEFT) AND
THE TRANSFORMED VARIABLE "MASS LESS 5MM IN G/TONNE" (RIGHT). .......................... 125

FIGURE 39: CROSS SECTIONS OF THE OREBODY AND HISTOGRAMS FOR THE THREE VARIABLES
ESTIMATED INDEPENDENTLY INTO A 0.7M X0.7M X 0.7M GRID AND ACCUMULATED INTO A 5M
X 5M X 5M GRID. ......................................................................................................................... 126

FIGURE 40: HISTOGRAMS SHOWING TRANSFORM OF DATA FROM RAW DATA TO GAUSSIAN
VARIABLES. .................................................................................................................................. 127

Stephen Coward - April 2020 17

 Evaluation of metallurgical recovery factors for diamonds recovered from kimberlites

FIGURE 41: VARIOGRAM MODELS (TOP LEFT AND BOTTOM RIGHT)AND CROSS VARIOGRAM MODELS
(BOTTOM LEFT) FOR THE GAUSSIAN TRANSFORMS OF P-WAVE AND DROP WEIGHT TEST DATA.
......................................................................................................................................................127

FIGURE 42: PERSPECTIVE PLOTS SHOWING A CROSS-SECTION THROUGH THE TEST AREA FOR 5M X
5M X 5M BLOCKS FOR SIMULATED AND KRIGED P-WAVE VELOCITY (UPPER IMAGES) AND DROP
WEIGHT TEST DATA(LOWER IMAGES). ......................................................................................128

FIGURE 43: SCHEMATIC OF THE OPTIONAL ROUTES TO USE POINT SCALE SAMPLE DATA TO PREDICT
THROUGHPUT. ..............................................................................................................................130

FIGURE 44: A RELATIONSHIP BETWEEN INPUT ENERGY AND DEGREE OF FRACTURE, EXPRESSED AS
PERCENTAGE PASSING 1/10TH OF ORIGINAL PARTICLE SIZE (BLUE DIAMONDS), SHOWING A
FITTED BREAKAGE FUNCTION IN BLACK. ...................................................................................131

FIGURE 45: PLOT SHOWING BROAD CORRELATION BETWEEN AVERAGE LONG RUN ENERGY
CONSUMPTION IN SEMI AUTOGENOUS GRINDING (SAG) MILLS AND AVERAGE OREBODY A*B
VALUES (DANIEL, LANE AND MCLEAN, 2010). ......................................................................132

FIGURE 46: RELATIONSHIP BETWEEN POWER CONSUMPTION AND THROUGHPUT FOR A TARGET
GRIND SIZE. ..................................................................................................................................133

FIGURE 47: HISTOGRAMS OF THE VARIABLES THAT ARE CALCULATED AND ESTIMATED IN PATHWAY
1. ...................................................................................................................................................137

FIGURE 48: HISTOGRAMS OF THE VARIABLES THAT ARE CALCULATED AND ESTIMATED IN PATHWAY
3, DATA HAS BEEN GROUPED BY PREDICTED THROUGHPUT QUARTILES. ...............................137

FIGURE 49: PLOT SHOWING THE VARIABILITY IN WEEKLY THROUGHPUT AND A TABLE WITH
SUMMARY STATISTICS FOR EACH OF THE VARIABLES CALCULATED THROUGH EACH PATHWAY .
......................................................................................................................................................138

FIGURE 50: A SCHEMATIC DEPICTION OF APPROACHES TO ESTIMATING THE RECOVERY FACTOR FOR
DIFFERENT PROJECT MATURITIES. .............................................................................................143

FIGURE 51: A PLOT OF THE SPECIFIC INPUT ENERGY AND CUMULATIVE PROBABILITY OF FAILURE
FOR A FEW SELECTED MINERALS, SIZE PARAMETER SET TO 5MM (ADAPTED AFTER KING
2001). .........................................................................................................................................158

FIGURE 52: A PLOT OF THE MEDIAN FRACTURE ENERGY FOR MINERAL PARTICLES OF DIFFERENT
SIZES. ............................................................................................................................................159

18 Stephen Coward - April 2020

 Chapter 1: Introduction

FIGURE 53: A PLOT OF ENERGY INPUT VS PRODUCT SIZE USING THE T10 APPROACH AND A ROSIN
RAMMLER BREAKAGE FUNCTION MODIFIED AFTER KING (2001). ...................................... 167

FIGURE 54: A DENSIMETRIC DISTRIBUTION PLOTTED FOR FOUR SAMPLES DERIVED FROM
KIMBERLITE, CRUSHED TO 100% PASSING 12MM AND GROUPED BY DENSITY CLASSES. ... 169

FIGURE 55: A PLOT OF THE LOG OF DIAMOND WEIGHT VS LOG OF THE NUMBER OF STONES IN EACH
CLASS PER HUNDRED TONNES PER UNIT INTERVAL. ................................................................ 171

FIGURE 56: A PLOT OF THE LOGARITHM OF DIAMOND SIZE VS STONE FREQUENCY IN STONES PER
HUNDRED TONNES PER UNIT INTERVAL. .................................................................................. 172

FIGURE 57: PLOT SHOWING THE ACTUAL AND MODELLED LOG NORMAL DIAMOND SIZE
DISTRIBUTION. ............................................................................................................................ 182

FIGURE 58: COMPARISON OF THE THICKNESS AND V1 BASE MAPS FOR THE KRIGED AND SIMULATED
OUTPUTS OF EACH SCENARIO WITH THAT OF THE V-BOD. GRADE WAS HELD CONSTANT FOR
EACH SCENARIO. .......................................................................................................................... 197

FIGURE 59: A DIAGRAM DEPICTING THE IMPLEMENTATION OF THE MINING CONSTRAINT LOGIC.
...................................................................................................................................................... 199

FIGURE 60: A DEPICTION OF LIBERATION AS A FUNCTION OF THE CONCENTRATION OF KIMBERLITE

IN THE HEADFEED. ...................................................................................................................... 200

FIGURE 61: A VIEW OF THE THREE LOBES OF THE DEPOSIT LOOKING FROM THE WEST TO THE EAST
(SOUTH LOBE IN DARK BLUE) ADAPTED AFTER CAMPBELL (2009). ................................. 207

FIGURE 62: SCHEMATIC OF PROCESS FLOWS USED FOR PROCESS MODEL. ..................................... 209

FIGURE 63: PLOT SHOWING A SIMULATION OF THE SIZE OF 20 000 DIAMONDS DRAWN. .......... 210

FIGURE 64: SCHEMATIC OF THE INTEGRATED EVALUATION MODEL ARCHITECTURE (ADAPTED

AFTER BURBECK, 1992) . ......................................................................................................... 211

FIGURE 65: SUMMARY OF TONNAGE TREATED ................................................................................. 214

FIGURE 66: SUMMARY PLOT OF CUMULATIVE DISCOUNTED CASHFLOW FOR THE AK PROJECT, P50
CASE SHOWN IN GREEN,P80 AND P20 CASE SHOWN IN RED, INDIVIDUAL CASES SHOWN IN
GREY. ............................................................................................................................................ 216

Stephen Coward - April 2020 19

 Evaluation of metallurgical recovery factors for diamonds recovered from kimberlites

FIGURE 67: GRADE SIZE PLOT SHOWING THE IMPACT OF APPLYING A STRICT SIZE CUT-OFF TO A
TOTAL CONTENT CURVE. .............................................................................................................217

FIGURE 68: A SCHEMATIC REPRESENTING THE PRIMARY RESPONSE FRAMEWORK FOR

GEOMETALLURGICAL VARIABLES. ..............................................................................................220

FIGURE 69: A DATA TYPOLOGY ADAPTED AFTER KEENEY AND WALTERS (2008). ....................221

FIGURE 70: A LANDSCAPE FOR SAMPLE TYPE CLASSIFICATION IN TERMS OF BOTH SPATIAL
CONTINUITY AND PRIMARY RESPONSE DIMENSIONS. ...............................................................222

FIGURE 71: SCHEMATIC SHOWING THE LOCATION OF SAMPLES FROM K2 PIPE. ...........................247

FIGURE 72: A VIEW OF THE LAYOUT OF THE CORE HOLES DEPICTING THE LOCATION OF THE
SUBSAMPLES ................................................................................................................................248

FIGURE 73:LISTING OF SUBSAMPLES TAKEN FROM CORES THAT WERE FULLY SAMPLED .............249

FIGURE 74: DEPICTION OF THE SUBSAMPLES TAKEN FROM CORES THAT WERE PARTIALLY SAMPLED.
......................................................................................................................................................250

FIGURE 75:A VIEW FROM THE SOUTH-WEST OF VENETIA K2 SAMPLED AREA - SHOWING
DOWNHOLE DENSITY. ..................................................................................................................251

FIGURE 76:A VIEW FROM THE SOUTH-WEST OF

VENETIA K2 SAMPLED AREA - SHOWING
DOWNHOLE DENSITY AND P WAVE VELOCITY. .........................................................................252

FIGURE 77:A VIEW FROM THE SOUTH-WEST OF VENETIA K2 SAMPLED AREA - UCS SAMPLE VALUES
AS SCALED SPHERES AND ESTIMATED BLOCK DENSITY IN TRANSPARENT BLOCKS. ..............252

FIGURE 78:A VIEW FROM THE SOUTH-WEST OF VENETIA K2 SAMPLED AREA - T10 SAMPLE VALUES
AS SPHERES AND ESTIMATED T10 IN TRANSPARENT BLOCKS. .................................................253

20 Stephen Coward - April 2020

 Chapter 1: Introduction

LIST OF APPENDICES

APPENDIX 1 -EXPERIMENTAL DESIGN AND DATA FROM OREBODY SAMPLING .............................. 247

APPENDIX 2 -OREBODY CHARACTERISTIC ESTIMATES .................................................................... 251

Stephen Coward - April 2020 21

 Evaluation of metallurgical recovery factors for diamonds recovered from kimberlites

1 INTRODUCTION

The specific limitations of traditional recovery factor estimation processes

addressed in this thesis are those that arise from the difficulty of incorporating
variable and uncertain rock properties in the derivation of the metallurgical
recovery factor. Failure to explicitly include the impacts of rock properties in
traditional project evaluation can result in sub-optimal mine design, incorrect
process configuration (design, operating strategy) and potentially biased
production and cashflow forecasts.

The metallurgical recovery factor is introduced in the context of the diamond mining
industry and its impact on diamond mining project evaluation. The connection
between the unique characteristics of diamonds (their particulate nature, low
concentration and logarithmic relationship between size and value), their host
deposits (kimberlite), and the diamond recovery processes is explained.

The consequence of using global estimates for rock and diamond characteristics,
based on few and spatially sparse data combined with the vast differences between
the scale of measurements from small samples and the scale of estimation of these
properties is described.

The primary response framework (Coward et al., 2009) is introduced to clarify the
taxonomy for variables that are used in this research. This conceptual framework
provides a basis for developing quantitative models of the relationships between
variables that drive uncertainty in the recovery of diamonds. These models are used
to propagate variability, and uncertainty through the diamond recovery value chain.
The specific uncertainties explored include those associated with data collection,
characteristic estimation, and process response modelling.

The introduction concludes with a demonstration of the role that the metallurgical
recovery factor plays in evaluating the economic potential of diamond mines.

22 Stephen Coward - April 2020

 Chapter 1: Introduction

1.1 Current approaches used to evaluate diamond mining projects

Evaluating the economic and technical viability of diamond mining projects requires
several inputs and assumptions. The report required by Canadian National
Instrument NI-43-101 (NI43-101, 2001) for standards of disclosure for minerals
projects comprises a detailed list of information and technical data vouched for by
a Competent (Qualified) Person. Many orebody inputs (e.g., grade, tonnage
estimates) are derived from limited, and/or spatially sparse, data and are thus
characterised by a high degree of uncertainty. One of these uncertain inputs is the
metallurgical recovery factor that is used to calculate the quantity and quality of the
diamonds that will be delivered by the project.

Current reviewed methods of estimating and using the metallurgical recovery factor
do not explicitly account for the impact that variable and uncertain kimberlite
characteristics have on the derivation and use of the factor. Assumptions of
continuity of kimberlite characteristics in large parcels of mined material has
potential to under- or over-estimate recoveries and in some cases, lead to material
inaccuracy in the estimate of project value (Mackey and Nesset, 2003).

This research project focusses on the development of a methodology that will

improve the evaluation process by incorporating both variability and uncertainty of
diamond and kimberlite characteristics in the evaluation of diamond mining
projects. The approach is centred on the development of a value chain model that
facilitates the propagation of variable physical characteristics of kimberlite through
the value chain from the orebody to final product. The value chain approach allows
the modelled variability of kimberlite characteristics to interact with constraints in
the mining and processing of kimberlite. The system model is used to quantify the
influence of the characteristics on process throughput, process efficiency and
diamond recovery. Historic methods have, for various reasons, not used this
approach.

The approach makes it possible to determine, at various scales of resolution, the

expected differences between the estimated in situ diamond population and the
population of diamonds that will be recovered. Passing multiple simulated
realisations of the deposit’s kimberlite properties through the value chain allows
the generation of a range of plausible estimates for the metallurgical recovery factor.
This range can be used to derive confidence limits or probability threshold values
for the metallurgical recovery factor over the life of the mining operation.

The potential value of a diamond project is driven, to a large degree, by the size of
the deposit, the in situ diamond grade and the in situ diamond value from the in situ
resource. Translation of the estimate of in situ grades of the orebody into an

Stephen Coward - April 2020 23

 Evaluation of metallurgical recovery factors for diamonds recovered from kimberlites

expected cash flow model for valuation of the mining project requires valid methods
to determine diamond recovery and loss . The metallurgical recovery factor is one
of the modifying factors used in the development of the valuation model. Its value
reflects the proportion of the total population of in situ diamonds that is expected
to be recovered when the mine is operational. This expected carat recovery is then
integrated with the other estimates into a financial model to estimate a mining
project’s value.

This evaluation process is a component of the mining value chain. A schematic of

this process is show in Figure 1.

Figure 1: A schematic overview of a mining value chain, and a breakdown of the

metallurgical recovery factor. Adapted after (Appleyard, 2001).

Development of a mining project begins with the discovery of a kimberlite that has
the potential to contain diamonds. Once discovered, several sampling and
measuring processes are initiated to gather data on the diamond content and
information that is used to infer the size of the deposit. These activities include
collection of geological information based on outcrops, core and trench samples.
The uncertainty associated with the grade estimates of portions of the resource is
used to classify the resource into inferred, indicated or measured categories as per
the requirements of the prevailing codes for reporting exploration results, mineral
resources and ore reserves, e.g. JORC (JORC, 2012) and SAMREC (SAMREC, 2016).

These resources are then subjected to a planning process that is usually divided into
several phases. At each phase the number of options, or configurations considered,
are reduced and ideally, as more information is acquired, confidence in the
estimates of the primary parameters (Grade, Tonnage, Deposit shape) of the mineral
project is increased. Once a mine plan and a proposed process plant design has been

24 Stephen Coward - April 2020

 Chapter 1: Introduction

completed, it is possible to estimate the expected tonnage and grade that will be
produced by the project. This calculation includes assumptions about both mining
and metallurgical process efficiencies. These efficiencies are used to estimate the
reserves of the project that can be classified as either probable or proven as defined
in the relevant classification code e.g., JORC (2012) code. The declared reserves are
an essential input into the mine and process plant design and configuration.

The evaluation process is iterative in nature, and usually includes several phases of
sampling, sample processing, resource estimation, mine and process design, reserve
estimation and valuation. In each iteration, more information on the deposit is
acquired, reducing uncertainty in the geological models, grade estimates, and other
inputs used to evaluate the viability of the mining project.

At the end of each assessment stage a gating process is carried out to select the ‘best
next step’ for the project. The next steps might include divesting from the project,
halting the project, continuing the original plan, or increasing expenditure to
expedite the project delivery. These decisions are informed by both the valuation
of the project and, to some degree, the uncertainty of the valuation. (Brennan &
Schwartz 1985; Bratvold & Begg 2002).

1.2 Specific challenges faced in evaluating diamond-mining

projects

The evaluation of kimberlite diamond deposits faces several specific challenges,

these include:

• A low proportion of kimberlite deposits contain diamonds;

• Kimberlites exhibit a range of geometries that result from multiple
phases of volcanic intrusion;
• The concentration of diamond mineralisation is very low;
• The mineralisation is of a particulate nature; and
• Diamond value is a function of several characteristics that are difficult
to sample in a representative way using small samples.

Low proportion of pipes are economic

Few of the kimberlites discovered to date contain sufficient diamonds to support a

mining operation. There are approximately 5000 kimberlite pipes that have been
discovered, of which around 50 have been mined at some stage, and about 15 are
major producers (Kjaarsgaard, 1995). However, approximately 58% by value of
diamonds mined are estimated to come from kimberlite pipes (De Beers Mineral

Stephen Coward - April 2020 25

 Evaluation of metallurgical recovery factors for diamonds recovered from kimberlites

Resource Intelligence Unit, 2005). Although diamonds also occur in other deposit
styles, such as alluvial and fluvial, this research is focused on kimberlitic deposits.

Kimberlite deposits exhibit a wide range of geometries

Kimberlites are the transport mechanism that brings diamonds from the diamond
stability field in the earth’s mantle into the crust in a medium that preserves the
diamonds (Field & Scott Smith 1999).They are formed by a diverse range of natural
volcanic events. Such events give rise to deposits that exhibit a wide range of
geometries and thus require substantial geological investigation to understand the
geometry of the phases of the deposit and the presence and dispersion of diamonds
contained within each phase (Field et al. 2008).

Economic diamond concentrations are very low

The grades of diamond deposits are typically expressed as carats per hundred
metric tonnes (CPHT). A carat is equivalent to 0.2g and thus this measurement
equates to a weight-by-weight fraction of one part per 500 million. Table 1 gives a
few examples of grades of mines that are, or were, operating at the time of
publication.

Average
Average Revenue
Mine production $/tonne Year
$/ct.
Grade in cpht.

Jwaneng 198 150 297 2016

Venetia 58 187 108 2013

Finsch 56 108 60 2018

Orapa 81 97 79 2016

Premier Mine 39 125 49 2018

Koffiefontein 6.4 525 34 2018

Williamson 7.6 270 21 2018

Table 1: Selection of several diamond mines showing the mine grade recovered in cpht,
average diamond value in $/ct and the average revenue in $/tonne ( Petra Diamonds,
2019; Anglo American, 2018; De Beers Group services 2018 )

26 Stephen Coward - April 2020

 Chapter 1: Introduction

Diamonds are particles with a wide distribution of sizes

Given the unique nature of diamond mineralisation, it is useful to consider the

methods and variables used to describe diamond content and the methods used to
analyse relationships between concentration and size. The variables include several
forms of the diamond mass concentration and diamond stone size distribution.

Diamonds exist as individual particles embedded in a kimberlite matrix. Unlike

assaying for metals where relatively small samples can be crushed and assayed to
determine the average grade of the contained metal, the diamond sampling process
requires the liberation and recovery of a mixed population of particles. The
particles of interest (diamonds) range in size from 0.1µm to over 100mm. The
abundance, by number, of small diamonds is higher than that of coarse diamonds.
Table 2 gives an example of a parcel of diamonds that has been sieved and the
weights retained on each sieve have been recorded.

Cts
Lower Unit
Diamond Average retained Stone
Critical interval
Sieve Size on each Count
Size Factor
screen
# Ct/Stone Ct/stone Ct #
+15 CTS 14.8 17.118 0.00 0
+23 8.0360 10.9060 3.7704 0.00 0
+21 3.6910 4.8500 2.9595 0.00 0
+19 1.9180 2.4800 3.5175 0.00 0
+17 1.4230 1.5700 7.7134 0.00 0
+15 1.1950 1.2600 13.1862 1.75 1
+13 0.7030 0.8600 4.3400 3.22 4
+12 0.5230 0.5610 7.7849 2.57 5
+11 0.3170 0.3710 4.5989 0.92 2
+9 0.1790 0.2110 4.0289 0.92 4
+7 0.1170 0.1230 5.4151 3.38 27
+6 0.0792 0.0896 5.9011 0.90 10
+5 0.0485 0.0730 4.6952 0.90 12
+3 0.0256 0.0350 3.6036 0.60 17
+2 0.0138 0.0210 3.7263 0.43 20
+1 0.0054 0.0140 2.4541 0.02 2
-1 0.0020 0.0090 2.3182 19.00 2111
Table 2: Analysis of diamonds sized by screening (adapted after Ferreira 2013)

Diamond sieves are metal plates with round holes of a specific diameter punched in
them. The diamond trading company nomenclature for the sieves is given in the
first column of Table 2. Each aperture size can be related to a characteristic
diamond weight that has a 50% probability of either passing through or being

Stephen Coward - April 2020 27

 Evaluation of metallurgical recovery factors for diamonds recovered from kimberlites

retained on the screen of this aperture; this is referred to as the critical size for that
screen aperture.

As the difference in the size of the orifices in the screen sequence varies between
sieve sizes, the masses retained on each sieve need to be factored or standardised
by the relative 'distance' of the interval between each sieve in order to compare
relative abundance in each size fraction. This factor, known as the unit interval
factor, can be derived from Equation 1

𝑼𝒏𝒊𝒕 𝑰𝒏𝒕𝒆𝒓𝒗𝒂𝒍 = 𝟏⁄
[𝐥𝐨𝐠(𝑪𝒔𝒛𝒖𝒑 ) − 𝐥𝐨𝐠 (𝑪𝒔𝒛𝒍𝒐𝒘 )] Equation 1
where 𝐶𝑠𝑧𝑢𝑝 is the critical size in ct/stone of the previous sieve used in the
sequence; and 𝐶𝑠𝑧𝑙𝑜𝑤 is the critical size in ct/stone of the sieve on which the
diamonds are retained. This unit interval is used to normalise the mass or number
of stones retained per sieve class when plotting various charts of the diamond size
distribution.

The diamonds retained in any sieve class can be reported as either a mass of
diamond or a stone count. The average size of stone that is retained on a sieve size
in the sequence is derived from sieving and counting stones retained on each sieve,
and then dividing the total mass of the stones retained on the sieve by the stone
count. This average can be used to generate an approximate estimate the number
of stones retained in a sieve interval for other sieved parcels of diamonds.

The information in Table 2 is also depicted in Figure 2, which shows diamond weight
class on the horizontal axis and diamond count on the vertical axis. The figure on
the left demonstrates the relationship between size and stone abundance. In the
plot on the right-hand side is calculated by taking the logarithm of the class average
stone size and applying the unit interval correction to the stone count in each class.

Diamond Size vs Stone count per Diamond Size vs Stone count per
weight class weight class with UI correction
500 120
450
Stone Count Per class
Stone Count Per class

400 100
350 80
300
250 60
200
40
150
100 20
50
0 0
0 5 10 15 0.01 0.1 1 10 100
Size Class (Average size Ct/stone) Log Size Class (Average size Ct/stone)

28 Stephen Coward - April 2020

 Chapter 1: Introduction

Figure 2: Plot of diamond stone counts per size class.

Diamond grade sampling is challenging

The purpose of the sampling process is to recover, as far as is possible, a good

representation of the full size spectrum of diamonds contained within the target
deposit. The relative abundance of smalls stones, and relative scarcity of large
stones, suggests that small samples will be biased, containing a higher proportion of
small stones and their average stone size will be below that of the true in situ value.
Samples of sufficient support are required to so that both the sampled stone
concentration as well as proportion of stones in each size class can be considered
to be sampled in a representative manner. It is also important to note that not all
the small diamonds will be recovered during the sample treatment process; those
smaller than the bottom screen aperture will be lost, and a number will also remain
‘locked’ in the coarse discard stream. Figure 3 has been generated by simulating
sampling of diamonds from the parcel shown in Table 1.

In each sampling iteration, the number of diamonds sampled is increased. The

green line shows how the average stone size estimated from ten samples of the same
number of stones increases with an increase in the size of the sample. The increasing
size of the sample also reduces the range between the maximum and minimum
estimate of the average stone size.

The graph also shows how the actual average stone size in the parcel is
underestimated (biased) because of the under-representation of larger stones in the
small samples. This demonstrates why large samples are required and even with
very large samples there is always a probability that the recovered diamond size
distribution may not be representative of the in situ size distribution. This is the
reason that models of the relative abundance in each size fraction are often used to
predict the diamond size distribution that will be recovered by a full-size production
recovery process.

Stephen Coward - April 2020 29

 Evaluation of metallurgical recovery factors for diamonds recovered from kimberlites

Sample Size vs Sampled Average Stone Size

0.4
Sampled Average Stone Size

0.35
0.3
0.25
(Ct/stone)

0.2
0.15
0.1
0.05
0
0 50 100 150 200
Number of Stones Sampled
Minimum Average Maximum Actual Average Stone Size

Figure 3: Variation of the average stone size estimated from small samples.

Relationship between diamond size and value

Diamond value is positively correlated with size. This relationship (see Figure 4) can
be described using a logarithmic model for a given combination of shape, colour and
quality. An accurate model of the in situ $/ct for each diamond size class in the
population is required to estimate the in situ $/tonne.

Figure 4: Typical relationship between diamond size and cut diamond retail value (adapted
from Rappaport, 2009).

If the parcel in Table 2 was valued using a $/ct per sieve class like that depicted in
Figure 4, its ‘bench value’ would be 140 $/ct. If the same parcel was sampled as

30 Stephen Coward - April 2020

 Chapter 1: Introduction

described above using samples containing 100 stones, the change in the sampled
size distribution produces samples with bench values between 471 $/ct and 46 $/ct.
The under-sampling, and low recovery, of large stones in small samples is a
substantial source of uncertainty in the evaluation of kimberlite deposits. Methods
that have been developed to address this are discussed further in the literature
review section and their impact is described in case study 2.

1.3 Diamond project evaluation practice

Given the cost of diamond sampling and the low probability that a randomly
selected kimberlite will have an economic diamond content, initial evaluation
activities following discovery are used to limit expenditure to that which is sufficient
to generate just enough data to decide whether the deposit is likely to contain
diamonds in quantities that will support economic extraction.

The assessment of several geochemical and petrological characteristics of surface

and core samples taken from the discovered deposit can be used to indicate the
potential for diamond content (Cookenboo & Grutter 2010; Macnae 1995). Should
the results indicate that the kimberlite is likely to be diamondiferous several cores
are extracted from the deposit and are then dissolved to recover what are termed
micro diamonds.

Micro diamonds are diamonds smaller than 0.5 mm in size that are recovered from
core through a thermo-chemical dissolution process. Typically, samples of the order
of 20 to 50kg are required. The analysis of micro diamond results is used to
determine the size frequency relationships that exist in the lithologies within the
deposit. This size vs stone frequency relationship is modelled and used to generate
an estimate of the diamond potential in the deposit by extrapolating the size and
stone frequency relationship from micro diamond sizes into larger diamond sizes.
(Rombouts, 1995)

Should the potential diamond content be deemed sufficient for investment purposes
then the next phase of evaluation is initiated, and usually aims to establish an
estimate of the financial value of the diamonds in the deposit. The financial value of
diamonds is influenced by several factors. These factors include diamond size, their
quality (for example number of inclusions, flaws), their colour, and their shape. The
valuation process starts with the definition of the size frequency of the diamonds.
As the number of diamonds acquired increases, the assortment of the diamonds in
the deposit begins to be revealed. The assortment refers to the shape, colour and
quality of the diamonds in a deposit. To achieve reliable estimates of diamond
values typically requires approximately three thousand to five thousand carats for
assessment (Rombouts, 1995). Figure 5 based on the sample depicted in Table 2

Stephen Coward - April 2020 31

 Evaluation of metallurgical recovery factors for diamonds recovered from kimberlites

shows a comparison between the relative proportion of carats and their

proportional contribution to value.

100% To acquire a sufficient

>1.5 ct
90% 10% number of carats
0.5-1.5 ct requires samples of the
80% 93%
order of 10 to 100
>1.5 ct
70%
50% tonnes each. The
60% sample size and the
50% numbers of samples
40%
required, however,
<0.5 ct
93% depend on both the
30% 0.5-1.5 ct
30%
estimated average
20%
grade derived from
10%
<1ct
micro diamonds as well
0% 82% as the geological
% of Carats % of Revenue heterogeneity of the
deposit.

Figure 5: Comparison of the proportion of carats and The assay, or sampling

proportion of revenue represented by different diamond size
for macro diamonds is a
ranges, for the diamond parcel described in table 2.
process that requires
large samples and
produces results that, to some extent, depend on the efficiency of the recovery
process. Any change in the efficiency of sample acquisition, sample crushing or
density separation process will change the number of diamonds that are recovered
from the samples.

Even with the time, cost and logistical complexity associated with the collection and
treatment of large samples, it is possible, with sufficient core and large diameter
drilling, to estimate the in situ macro diamond grade with reasonable confidence. It
has been demonstrated that diamond grades are spatially correlated variables that
can be estimated using spatial estimation techniques such as kriging. (Kleingeld,
1996; Rombouts, 1995).

32 Stephen Coward - April 2020

 Chapter 1: Introduction

1.4 Use of the metallurgical recovery factor in project evaluation

The metallurgical recovery factor is an expression that is used to convert the

estimated in situ grade into an expected recovered grade. At an early stage of the
project, the factor is expressed as a proportion of the recovered diamond grade
divided by the estimated in situ diamond grade, in either a mass proportion (CPHT)
or stone proportion (SPHT). In the absence of detailed spatial data, the recovery
factor is derived heuristically (e.g., by comparison to analogous deposits of similar
size and composition) and assumed to be a constant that is applied globally across
the entire deposit.

As the project matures, and more information on the project is acquired, it becomes
possible to improve the methods used to derive the recovery factor and increase the
spatial resolution of the recovery factor, both into smaller areas of the deposit
and/or to shorter periods of production that are associated with the mining and
treatment of specific areas of the deposit. The form of the factor can also be adapted
to express the expected recovery per diamond size fraction. The recovery factor
does not depend on a single variable tied directly to a regionalised variable, but is
the result of the complex interaction of the:

• mining and metallurgical processes used in diamond liberation, separation

and recovery;
• mined country rock and kimberlite characteristics; and
• characteristics of the diamonds in the deposit.

These relationships are briefly described in the following sections.

1.5 Impact of recovery process design and operation on

metallurgical recovery.

The recovery process aims, as far as is possible, to preserve the in situ diamond size
distribution between the predetermined upper and lower size limits of the process.
To do this the ore is crushed in stages with several cycles of diamond removal from
the crushed products. The process flow sheet can be considered as having three
primary objectives;

1 - Comminution of the kimberlite to liberate the diamonds,

2 - Separation of the diamonds from the gangue using density, and

3 - Recovery of the diamonds from the dense concentrate.

Stephen Coward - April 2020 33

 Evaluation of metallurgical recovery factors for diamonds recovered from kimberlites

To achieve this efficiently several different size streams are generated by screening
and each size stream is then subjected to slightly different processes that are
tailored to the size distribution that is to be treated.

The selection of design parameters for the treatment of a given kimberlite deposit
is based on a combination of deposit scale, assumptions about the kimberlite
characteristics and long run performance of the selected unit processes. During the
early stages of the project the recoverable diamond size envelope is determined,
which defines the top and bottom cut-off size of diamond that will be recovered. As
the project progresses several trade-offs studies are undertaken with the aim of
optimising the balance between increased recovery and increased capital and
operating cost of selected treatment processes.

If the host rock is crushed very finely a larger number of diamonds will be released
but there will also be an increased probability of diamond damage and or breakage.
The range of recovered diamond size and the sequence of crushing and separation
that is used are very important considerations in the design and management of the
operation. The configuration of these processes also has a material impact on the
proportion and size distribution of diamonds that are liberated, recovered and lost
through lock up and damage.

Blasting, which is the first phase of comminution, is a process used to extract the
rock and reduce its size for delivery to the recovery plant. This changes the
properties of the rocks to some extent and causes some diamond damage. Studies
have, however, shown this to be negligible beyond more than five blast hole
diameters (Wilmott, 2004).

In-pit and primary crushing usually produce a product with a top size that is in the
range of 250 to 125mm. This reduces the feed to a manageable size for entry to the
plant. The next stage of comminution reduces the material to a nominal cut size of
32mm usually through a combination of cone and impact crushers.

The material is then washed, and fines removed in preparation for dense media
separation. In the Dense Media Separation (DMS) process the diamond particles,
having a density of 3.5 g/cm3, are separated from the particles of kimberlite, most
of which have a density of less than that of diamond. Several machines are used to
do this, but by far the most common is the hydrocyclone. This unit processes a
mixture of water and ferrosilicon (the dense media) into which the crushed ore is
mixed. This mixture is pumped into a hydro-cyclone and the dense particles pass to
the spigot or underflow whilst the less dense particles migrate to the vortex finder
and hence into the overflow.

34 Stephen Coward - April 2020

 Chapter 1: Introduction

The tailings from this process may contain diamonds in two forms:

1- Diamonds that have been incorrectly classified due to operational errors

2- Diamonds that have been correctly classified but have floated out of the
separation process because the combined density of the particle consisting of
diamond and kimberlite is less than the effective cut-point in the dense media
separation process.

This stream may be re-crushed to liberate these so-called ‘locked’ diamonds and to
recover any errant ‘free’ diamonds.

The undersize material and coarse oversize material from the process are usually
disposed in so-called ‘slimes dams’ or ‘processed kimberlite dumps.’ Due to
changing market conditions, and improving technology, over time these dumps may
become economic to exploit.

1.6 Impacts of kimberlite properties on treatment process

efficiency.

The impact that rock properties have on the process depends on the interaction of
the rock properties with the process. This interaction is controlled by the following
variables:

• Variables associated with the process design and configuration;

• Variables associated with the process operation,
• Variables associated with gangue that is mined with the kimberlite.

Kimberlites have been described, from a processing point of view, as a relatively soft
clay-rich rock, however due to the process of emplacement they may contain
differing amounts of several rock types that are mixed into the kimberlite (Boychuk
et al. 2012). The process of mixing, and the contents of the resulting mixtures means
that kimberlites comprise a wide range of rock types that have numerous and
variable physical characteristics.

Impacts of kimberlite properties on comminution

Kimberlite contains clay minerals (montmorillonite, smectite etc.) and, when

exposed to moisture, these clays swell and can materially change the strength of the
rock that is processed (Boshoff et al., 2006). Moderate clay content reduces the
strength of the rock and makes it more amenable to crushing, increasing crusher

Stephen Coward - April 2020 35

 Evaluation of metallurgical recovery factors for diamonds recovered from kimberlites

throughput and reducing the average size of the particles produced. Excessive clay
content can cause the particles to become sticky, cause agglomeration in the
crushing chamber and reduce throughput due to clogging of crushers, feed and
discharge chutes.

Fine grained fresh or hypabyssal kimberlites exhibit high rock strengths (e.g.,
Cullinan Mines Hypabyssal kimberlite has been measured to have a uniaxial
compressive strength of 150Mpa). At this hardness, the energy required to crush
this material can exceed the installed power of the comminution circuit. This can
lead either to a reduction in the overall grind of the material or, for a given grind,
require a reduction in the plant throughput.

Hardness variability can be difficult to deal with as the mass balance in the
comminution circuit can vary to an extent where one part of the circuit is
overloaded, which reduces overall throughput. Although this can be dealt with by
increasing stockpile sizes, this is not always a feasible option, and during overload
periods these reduced throughputs may not be detected if they are of short duration.

Impact of kimberlite properties on separation

The primary variable that controls separation is the apparent density distribution
of the rock for a given rock size distribution. Species denser than the effective cut-
point will sink along with the diamonds and species that are less dense will float and
either be discarded or re-crushed to release more diamonds. To measure the
proportion of material that is above the density cut point for a given lithology the
samples are crushed to a nominal size. These are then put through a sequence of
density separations in a static bath containing fluids of differing density. This
distribution of density is used to predict the mass flows across the process. It also
predicts the number of diamonds that are expected to be recovered from re-
crushing the floats from the first pass DMS process. Depending on the mineralogy
of the kimberlite there may be sizes of clasts that, when liberated, make a substantial
difference to the density distribution. An example of this is the modal sizes of
garnets that are dense and report to recovery.

Clay and fines content of the kimberlite may compromise the DMS operation by
causing turbulence in the cyclone. Surges of high density or incorrectly sized
material may cause overloading of the cyclone and reduce the rate of settling in the
cyclone. Flat particles generated from brittle rock types also hinder the efficiency
of the separation process. (Plitt, 1976)

36 Stephen Coward - April 2020

 Chapter 1: Introduction

Impact of kimberlite properties on final recovery processes

Final recovery processes use several diamond properties to reduce the mass of the
concentrate from the DMS and upgrade the final product that is dispatched from the
mine to diamond trading companies. The most common of these processes is the
use of x-ray fluorescence to separate the diamonds from the concentrate. Kimberlite
can, however, contain several minerals that fluoresce in a similar way, which can
compromise recovery efficiency. Magnetic separation is also used to reduce the
magnetite and other magnetic and paramagnetic particles from the recovered
concentrates. One of the final processes used on mine sites is hand sorting where
trained pickers are used to separate gangue from a rich diamond concentrate. Prior
to valuation recovered diamonds are washed in various acids to remove any
coatings that may impair downstream diamond grading and classification.

1.7 Impact of diamond properties on the recovery factor

The nature of diamonds makes them amenable to what is a relatively simplistic

physical recovery process. The main challenge in the process is to liberate and
recover as many diamonds as possible and yet preserve their size and quality
distribution to maximise the revenue derived from the recovered diamonds. The
effectiveness of the process in achieving this aim will determine what proportion of
the total population of in situ diamonds will be recovered. A variety of recovery
factors can be used to describe the relationship between the in situ diamond
characteristics and the characteristics of the recovered diamonds.

The shape of the size distribution of diamonds will display some variation between
deposits but can in most cases be modelled with a log-normal distribution.
(Kleingeld et al, 1996). The crushing and screening process are normally designed
to recover a portion of this size distribution that yields the highest proportion of the
revenue. The diamonds above the top cut-size are crushed, and those below the
bottom cut-off size are discarded as tailings. Premier Mine has one of the largest top
cut-sizes at 65mm, more commonly however this is set to 25 to 32mm. As most of
the screening is carried out on screens with square or rectangular openings, there
is always some degree of misclassification. This misclassification impacts on both
the mass flow of material and the shape of the recovered diamond distribution.

The selection of these cut sizes for a given mine is based on the average size
distribution across the deposit. This is integrated with the revenue distribution to
arrive at an expected $/tonne that will be recovered. This value, less the cost of
treating the material, will determine the contribution that will be derived from each
tonne of ore. This value is used to design the ultimate pit, the sequence and schedule

Stephen Coward - April 2020 37

 Evaluation of metallurgical recovery factors for diamonds recovered from kimberlites

for mining. This highlights how the estimation of the metallurgical recovery factor
impacts on strategic decisions in mine design and evaluation.

Diamonds are formed in the deep mantle, and during the emplacement process may
be subjected to several cycles of heating, cooling and mechanical stresses. This may
build up to a point where the diamonds become stressed and will be damaged in the
comminution process. Little quantitative work has been undertaken to estimate the
proportion of diamonds that are exceptionally stressed in situ, although as the
diamond cutting and polishing industry has shown, stresses in recovered diamonds
can impact on the yield that can be obtained when cutting and polishing the stones
(Rombouts, 1995).

Several shape and size distribution classification systems have been developed for
diamond populations (Caers and Rombouts, 1996). Their cubic crystal form is a
face-centred cubic lattice but there are many crystal habits to which diamonds
conform. The shape distribution has an impact on the way in which diamonds
liberate from kimberlite, the way in which they settle in the DMS process and the
trajectories exhibited in both magnetic and X-Ray processes. Surface features also
impact on the way in which diamonds liberate. Incompletely liberated diamonds,
sometimes referred to as 'comets', may float out of the dense media separation
process as the adhering kimberlite is less dense than the diamond so that the
aggregate particle density is less than that of the dense medium. These composite
particles may also be lost in the X-ray recovery section as the kimberlite may shield
the diamond from the incident X-rays and prevent the diamond from luminescing
sufficiently to be detected. The fine ferrosilicon media used in the DMS may adhere
or enter cracks in the diamonds, and as this media is magnetic, diamonds
contaminated in this way will be lost in magnetic separation units that are used in
final diamond recovery plants.

Although the colour distribution of diamonds, caused by several impurities, does

not always impact on the diamonds X-ray induced fluorescence, the presence of
impurities in the diamond may give rise to fluorescence characteristics that impair
the effectiveness of X-ray recovery machines. The fluorescence characteristics that
impair recovery include the total brightness that the stone will achieve (total
luminescence), the time to achieve this brightness (rise time), and the duration of
residual luminescence.

1.8 The sources of uncertainty in estimating the recovery factor

Kimberlites contain heterogeneous mixtures of rocks. When mined, the properties

of the daily feed to the treatment plant will vary substantially as the feed will contain
a mixture of rocks from different locations. It would be costly and very time

38 Stephen Coward - April 2020

 Chapter 1: Introduction

consuming to generate the number of samples that would be required to be

representative of the full range of kimberlite characteristics within each lithology or
facies. The response of the process to each individual rock type may also be
different to the process response to a mixture of rock types. One approach used to
overcome these limitations is to make up composite samples of kimberlite material
sourced from separate lithologies and mixed in predetermined proportions. The
proportions are based on the expected mixtures that will arise during mining
operations. Results from tests carried out on these mixtures, or composite samples,
are used to design the process and estimate the process capability (Lewis, 2001;
Ashley and Callow, 2000). The assumptions implicit in this approach include:

• variations in rock properties are stable within a geological domain;

• a representative sample of all the rock and mineral combinations can
be acquired;
• samples can be composited in such a way as to represent worst, best
and average scenarios;
• response of composites at sample scale can be equated to responses
of full-scale operation;
• there is complete knowledge of the proportion and the range of
mineralisation;
• blending achieved during sampling will closely mimic that of mining;
and
• historically derived scale-up factors can be used to scale-up the
results of tests to estimate process response of a full-size unit process.

The approach does not deal explicitly with expected variation of the recovery factor.
It can be demonstrated that there is a squared relationship between the uncertainty
in the estimate of the recovery factor and the uncertainty in the recoverable grade.
(Lantuéjoul, 1990). Hence, if the uncertainty in the recovery factor is halved,
uncertainty in the recoverable grade will be quartered. Under-estimation of the
magnitude of the uncertainty in recovery can prevent projects from reaching their
full potential and, in extreme cases, lead to cessation of operations shortly after
starting up. Conversely, by improving the protocols and methodology to estimate
this factor and reduce the uncertainty, it is possible to increase substantially the
value that can be attributed to reserves. (Benicelli et al, 2000)

Acquiring data of rock characteristics with enough spatial coverage and support to
generate unbiased estimates is problematic. Even if sufficient data are obtained
from a high number of samples with large supports to generate unbiased rock
property estimates, the use of these estimates to predict process response at
production scale requires adaption of existing methods of modelling process
performance. One approach to address scale-up explicitly, termed ‘Comminution
Economic Evaluation Tool’ (CEET) developed by MinnoveEX, now owned by SGS, is
Stephen Coward - April 2020 39
 Evaluation of metallurgical recovery factors for diamonds recovered from kimberlites

a system developed for short-term optimisation. CEET addresses the scale-up issue
by adapting the parameters of the prediction model so that the outputs of the model
match actual production outputs (Ameluxen et al., 2001). This approach provides a
good way to modify the inputs into the model by conditioning the models with data
once the plant and mine are running, but it is not suitable for prospective operations,
and has limited forecasting power.

Current practice in De Beers makes use of a combination of core samples and bulk
sampling. This includes the measurement and testing of a range of rock
characteristics, as described in the literature review. Diamond liberation and lock-
up is predicted based on a combination of rock size reduction measured during bulk
sampling and process models based on average kimberlite characteristics. This
approach has several shortcomings including:

• sample selection is based on geological domains that have been

derived from grade continuity and not necessarily processing
characteristics;
• samples are consumed during testwork and hence usually must be
drilled specifically for estimating a single ore property (‘it's not
possible to break the same rock twice’);
• test methodologies aim to ensure that the parameters of the process
are kept within strict tolerances and, therefore, do not capture the
impact that the material being tested has on the characteristics of the
process;
• as kimberlite contains clays that rapidly alter the physical
characteristics of the rock if it is exposed to moisture, some test
results may be compromised by sample degradation. In addition, if
there is a variable time between blasting and treating, the measured
laboratory response will not be a valid input for predicting the
operational process response;
• using small-scale tests based on composite samples is problematic.
Measured results from mixtures are only applicable to the specific
mixture that is tested. If the proportions are changed the results may
vary in a non- linear fashion. Resulting data cannot be used to model
relationships between primary rock characteristics and specific
process responses for individual lithologies. It is often possible
mathematically to mix results from tests on specific rock types, but
invariably impossible to un-mix results derived from composite
sample treatment;
• timing of data - bulk samples are often only taken at the feasibility
stage and hence pilot plant process information is not available at
early stages in new projects requiring the recovery factor to be made

40 Stephen Coward - April 2020

 Chapter 1: Introduction

on data derived from core samples that have been used to model and
predict the in situ diamond distribution.

1.9 Research model

Although work has been done to improve the methods used for sampling, ranging
from drilling techniques to improved laboratory practices, the underlying problem
of integrating the data acquired from sampling and estimation of rock properties
with the process design over the life of the mine has not yet been addressed.

Against the discussed background information, the research led to a methodology

for recovery factor estimation that utilises an integrated value chain of the operation
as shown schematically in Figure 6.

A sampling strategy that collects and integrates both destructive and non-
destructive data from core sampling was designed and executed. The data from this
sampling experiment are used to estimate and simulate the physical rock
characteristics into an orebody model.

Integrated Value Chain Model Simulation

Kimberlite
Mining Process Output
Property
Model Model Model Analysis

Estimate and Develop and Estimate and

Simulate Parameterise Simulate
Kimberlite Transfer Kimberlite
Characteristics Functions Characteristics

Kimberlite Simplified Mining Derivation and Use

Sampling and and Metallurgical of Metallurgical
Measurement Process Modelling Recovery Factors

Research Areas

Figure 6: A schematic representation of the research areas covered in this thesis.

This orebody model is depleted and treated using simulated mining and
metallurgical processes to evaluate the impact that the spatially varying rock
properties have on the treatment plant processes. A quantitative diamond recovery
model that responds to the rock characteristic variability and includes uncertainty
was developed. The process model is used to simulate the operation and produces
data that can be used to estimate both the expected diamond recovery and the

Stephen Coward - April 2020 41

 Evaluation of metallurgical recovery factors for diamonds recovered from kimberlites

uncertainty in metallurgical recovery for different periods of the mine’s operation,

at different scales of resolution.

The relationship between the sources of the variances in the integrated model can
be formulated as in Equation 2:

 rf2 = F ( r2 )G( p2 ) H (Cov. r2 p2 ) Equation 2

Where:

 rf2 : Variance of the recovery factor

 r2 : Variance of the rock properties

 p2 : Variance of the recovery process efficiency

(Cov. r2 p2 ) : Covariance between the rock and process properties

F, G, H are functions that are derived from either sampling and modelling or
observed relationships between the rock characteristics and the specific
process that is being considered. They must be calibrated for each specific
facies of the kimberlite bodies, and to the specific unit processes.

The building of the integrated value chain model required the following objectives
to be met:

• Development of an understanding of the relationships between geological

descriptions of kimberlite characteristics (e.g., texture, grain size) and
sampling instrument responses and destructive test responses;
• Review of several non-destructive geophysical measuring techniques that
could potentially augment the destructive data that are collected;
• Model the relationship between measured (or estimated) rock
characteristics and processes of comminution, separation and recovery;
• Develop or apply an appropriate methodology to create a spatial
representation of rock properties,
• Develop or apply a methodology to simulate the mining and metallurgical
processes that will respond appropriately to the estimated and simulated
rock characteristics; and

42 Stephen Coward - April 2020

 Chapter 1: Introduction

• Develop an integrated model of the mine operation to quantify the impact of

the spatial and temporal correlations of rock characteristics on the
metallurgical recovery factor.

1.10 Thesis structure

The literature review (Chapter 2) that follows addresses the nature of the
kimberlite, sampling theory, and sampling methods developed in this research to
augment limited destructive data with geophysical data. Spatial estimation and
simulation techniques used to generate models of the physical characteristics of the
orebodies being mined is followed by a review of mining and metallurgical process
simulation techniques. The literature review concludes with a statement of the
research problem (Chapter 3).

The first area of investigative research (Chapter 4) covers data acquisition and
describes the design and execution of a rock property sampling experiment that was
conducted to generate data required for this research project.

The data gathered are used to explore techniques for using non-destructive data to
augment the costly destructive data and estimate spatial kimberlite rock
characteristics. (Chapters 5)

A second area of research (Chapter 6) compares several methods of adapting

traditional mineral process models so that they can be used with spatial orebody
models to derive the metallurgical recovery factor and assess the impact of
incorporating stochastic uncertainty in process models.

The integrated model has been developed and applied to several projects during
this research process. Two case studies are presented to demonstrate the
implications of model outcomes on real diamond projects (Chapters 7 and 8).

A brief discussion follows to describe how the outcomes of this research can be
applied in an industrial framework (Chapter 9). The thesis concludes in Chapter 9
with several recommendations for further research.

The research was conducted in the context of this background to improve the
estimation of rock and diamond characteristics, model the complex interaction of
these with the recovery processes, and predict process responses at the scale of
diamond operations. The findings of this research provide an important milestone
in the development of methods to improve the derivation and use of the
metallurgical recovery factor. This will result in improved evaluation, design and
operation of diamond mining projects.

Stephen Coward - April 2020 43

 Evaluation of metallurgical recovery factors for diamonds recovered from kimberlites

2 LITERATURE REVIEW

To understand the magnitude of the challenges faced in deriving metallurgical

recovery factors, current methods of sampling, measuring and spatially estimating
rock characteristics in kimberlite diamond deposits are investigated. The
limitations and constraints in using this data in models of the diamond extraction
process are described. The aim is to identify clearly how it might be possible to
improve the derivation and use of metallurgical recovery factors by integrating
spatial models of rock characteristics with metallurgical simulation models.

The review begins with the description of the current understanding of kimberlite
emplacement processes and the consequent composition of kimberlitic diamond
deposits that these give rise to. It is shown that kimberlites exhibit a wide range of
rock characteristics and that this heterogeneity must be considered when sampling
for, and estimating, the characteristics required to derive metallurgical recovery
factors.

The current state of the art in sample acquisition, sample measurement tools,
quality assurance and control techniques reviewed. Non-linear relationships
between primary rock characteristics and process response variables are explored,
as is the impact that so-called "non-additivity" has on spatial estimation, and up-
scaling (from laboratory test to block-scale response predictions) of the estimated
or predicted variables. If incorrectly addressed, this important aspect of recovery
factor estimation can lead to significant undetectable bias in estimates. Using the
correct samples, measurement tools and sampling strategies can assist by
determining the nature of these relationships that can then be accounted for in the
subsequent estimation or simulation of the spatial characteristics of interest.

A description of mathematical modelling of mineral processing practices is provided

with explicit focus on comminution and dense media separation (DMS) processes.
Their use in recovery evaluation and limitations, both technically and conceptually
are discussed. The aim is to identify ways in which models of these two processes,
which play an important role in diamond recovery, can be used in a value chain
model that explicitly connects, at a block scale, with estimated and simulated
orebody models.

The review demonstrates that current practical approaches used to propagate rock
characteristic variability through the mining value chain have been hampered by

44 Stephen Coward - April 2020

 Chapter 2: Literature review

both availability of suitable software tools and limited processing capacity. This
review describes the evolution of optimisation models used in mining projects. It
demonstrates how a single optimised solution may not result in robust project
design if there is any orebody or processing uncertainty. The review concludes with
a description of a conceptual framework for an Integrated Evaluation Model. This
model aims to provide more robust estimates of the metallurgical recovery factor
using an integrated mining project value chain simulation.

The benefits of this approach are described as are the philosophical and practical
challenges presented by adopting a systems approach to a complex stochastic
problem.

2.1 Physical characteristics of kimberlites and diamonds

Kimberlites are pipe-like bodies that carry diamonds, formed deep in the earth’s
mantle, to the surface of the earth’s crust (Field, Stiefenhofer et al., 2008).

The objective of reviewing the literature on kimberlites is to gain an understanding

of the range of physical properties that this rock type exhibits and to understand the
implications of these on sampling, property measurement, estimation and process
modelling and simulation.

The research is focussed on one mineral in one geological setting. However, when
considered in the spectrum of rocks, kimberlite is an extreme end-member. There
are numerous processes that give rise to these bodies both during and after
emplacement. Not only do kimberlites exhibit a diverse range of geometries, the
post emplacement processes, such as weathering and alteration, add further
complication to the rock and mineral assemblage. This makes the design and
implementation of sampling campaigns capable of achieving representative results
particularly challenging.

Important aspects of kimberlite geology that impact on the sampling measurement

and estimation of the physical characteristics of the rocks to be mined and treated
include:

• The origins of Kimberlite bodies and the nature of diamond mineralisation.

• Clear definition of the rock property measurement problem by describing
the range of rock types that need to be considered and implications for rock
property measurement.
• An investigation of the range and variation of physical properties within a
given geological domain and between domains. This leads into a discussion
on the implications for defining zones within which estimates are to be made.

Stephen Coward - April 2020 45

 Evaluation of metallurgical recovery factors for diamonds recovered from kimberlites

• Identification of the factors that govern rock strength and how these are
dealt with in other rock types. This includes measurements of rock texture,
fracture frequency and measures of weathering and alteration in similar rock
types.
• A comparison and contrast of geological classification systems, their
objectives and the implications this has for both sampling for rock property
and estimating rock properties at a larger scale.
• Understanding the implications of the great differences in the scale of
measurement and the scale of estimation. This includes an assessment of the
challenges that these large differences present for change of support
modelling especially when the models must deal with non-additive, highly
skewed data.
• The scale on which this research requires estimates of rock characteristics
and the implication for mineral processing and diamond recovery modelling
and simulation.
• Identifying and modelling relationships between qualitative and quantitative
measurements of rock characteristics

An overview of kimberlite geology

Kimberlites are the host rock of numerous diamond deposits. They are a suite of
rocks that show great diversity in morphology, composition and hence mechanical
properties. They usually form pipe-shaped deposits of volcanogenic origin that are
mixed with country rock fragments and exhibit variable degrees of alteration. The
differences in kimberlite composition between pipes and even within the same pipe
can be substantial, these differences are exhibited from microscopic to macroscopic
deposit scale. Due to the economic significance of these deposits, classification and
nomenclature used to describe them continues to receive considerable attention.
Kimberlites were formally named after their type location, "Kimberley", in 1887 by
Henry Carvill Lewis at a meeting of the British Association for the Advancement of
Science in Manchester. He described them as “porphyritic mica-bearing peridotite
and suggested that they be recognised as a type of volcanic breccia.” (Lewis
referenced in (Mitchell, 1996)

The concept of petrologic clans was introduced by Reginald Daly in 1914 (Mitchell,
1996) and is predicated on the assumption that it is possible to classify rock types
into several suites by interpreting the composition of the magmas that led to the
rock formation. To a large degree, the original descriptive schemes have been
adapted by several authors. These include Skinner and Clement (1979) who
developed a system based on the modal mineralogy of the groundmass, based on
the belief that the ubiquitous presence of olivine was insufficient grounds for
classification. Since then several schemes have been proposed and these are
reviewed in more detail in this section.

46 Stephen Coward - April 2020

 Chapter 2: Literature review

The challenges faced by any generic description system include the requirement to
describe the following characteristics across a diverse range of scales and
emplacement settings:

• location, extent and geometry of the deposit;

• the geometry and location of the contained lithologies and smaller scale
structures, such as faults and fractures;
• content of the rock (elemental, mineral composition and rock types);
• genesis and spatial relationships; and
• alteration since emplacement.

Kimberlite occurrences, distribution and architecture

1.0E+11
1.0E+10 Ryholites
1.0E+09
Magma Vsicocity (Pas)

1.0E+08 Dacites
1.0E+07
1.0E+06 Andesites
1.0E+05
1.0E+04 Basalt
1.0E+03
1.0E+02 Kimberlites
Komatiites
1.0E+01
1.0E+00
500 750 1000 1250 1500 1750
Temperature (Degrees Centigrade)

Figure 7: Melt viscosity as a function of temperature at 1 bar for natural melts spanning the
compositional range Rhyolite to komatiite. All compositions are volatile free. The
temperature range is illustrative of typical eruption temperatures for each composition –
adapted from (Sperra, 2000) and (Sparks et al., 2009).

In the range of styles of volcanism, kimberlites can be viewed as an end member

(see Figure 7). They form at low temperatures from ultra-basic, low viscosity
magmas that range in age from early Proterozoic to early tertiary. This low viscosity
and low temperature of emplacement has consequences for the expected pipe
geometry and the geometry of internal contacts in these deposits. They are very
commonly preserved as downward tapering pipes and craters that have been
infilled with clastogenic mixtures of kimberlite and country rock clasts (Sparks et
al., 2005).

Stephen Coward - April 2020 47

 Evaluation of metallurgical recovery factors for diamonds recovered from kimberlites

Figure 8: A schematic view of a kimberlite pipe after Hawthorne (1975).

The shape and size of the deposit is related to the nature of the rocks into which the
pipes have intruded. A schematic cross section of a typical kimberlite pipe geometry
is presented in Figure 8. At least three types of kimberlite pipes have been identified:

• Deep (up to 2 km), steep-sided pipes which comprise three distinctive zones
(crater, diatreme, root)
• Shallow pipes (<500m) which comprise only a crater zone and are in-filled
exclusively with volcaniclastic kimberlite, mainly pyroclastic material, and;
• Small (<600-700m deep), steep-sided pipes filled predominantly with re-
sedimented material and less common pyroclastic kimberlite or, in a few
instances, with hypabyssal kimberlite (Field and Smith, 1999).

Even though most of the deposits seen today have been eroded to an extent where
the products of the emplacement volcanism are not present, some of these products,

48 Stephen Coward - April 2020

 Chapter 2: Literature review

such as fine ash and pyroclastic surge deposits, can be found within the body of the
pipes.

Kimberlites also occur as sills and dykes, and similar kimberlite pipes can contain
highly variable grades (e.g. Marsfontein, South Africa). However, even if one only
considers the pipe-like bodies, the range of rock types that can be identified is
substantial. A breakdown of the proportional contents that might be found in a
kimberlite deposit is shown in Figure 9.

% of Mineralogical
Deposit type Component % of Deposit zone Rock Description Rock analysis Modal analysis analysis

50 Fresh K feldspar
10 Altered K feldspar
Tuff ring 10 30 Country rock 100 Granite Fragments 30 Quartz
Fragments 10 Biotite

70 Serpentine
30 Macro crysts 100 Olivine Xeno crysts 20 Calcite
10 Smectite

Crater 20 30 Olivine pheno crysts 100 Serpentine

90 Kimberlite
Pipe 30 Garnet Xeno crysts 90 Garnet
10 Pheno crysts
10 Kaliphyte

30 Ilmenite Xeno crysts 100 Imenite

Pipe body 60
(Diatreme) 60 Serpentine
20 Smectite clays
30 Matrix Microcrysts 5 Spinel
Root zone 10 5 Monticellite
5 Phlogopite
5 Calcite
External Country rock lithics > 65mm
10
Waste

Figure 9: A schematic representation of the materials that could be found in a ‘typical’ kimberlite
pipe, adapted after (Field and Smith, 1999).

Work by Sparks (2006) on several of the southern African pipes, has suggested
that the rock types in the diatreme could be classified as four basic types:

• Massive volcaniclastic kimberlite (MVK)

• Layered volcaniclastic kimberlite (LVK)
• Marginal wall rock breccias (MWRB)
• Hypabyssal kimberlite (HK)

The massive volcaniclastic kimberlite (MVK), and layered volcaniclastic kimberlite

(LVK) can be considered as having similar components. The primary components of
this type of rock are listed in Table 3.

The kimberlite minerals are dominated by various forms of olivine (phenocrysts,

macro-crysts and xenocrysts). A distinction can however be made between group
one and group two kimberlites, with group one kimberlites containing olivine micro

Stephen Coward - April 2020 49

 Evaluation of metallurgical recovery factors for diamonds recovered from kimberlites

phenocrysts (groundmass perovskite, spinel, and monticellite or calcite) and group

two kimberlites, which have phlogopite as a major mineral.

Component Examples
Juvenile components derived from the magma Lapilli, phenocrysts and groundmass minerals
Minerals, xenocrysts Xenocrysts related to the breakup of mantle
nodules or basement crystalline rocks
A mega-cryst suite Crysts including olivine ilmenite and garnet
Xenoliths of deep origin Peridotite and eclogite
Country rock accidental lithics Can be correlated with the geology of the
basement and host country rock
Table 3: A description of the components of volcaniclastic kimberlites.

Layered volcaniclastic rocks can be distinguished from the massive type by visual
inspection of the large-scale rock fabric. There are many sub-facies in this category
reflecting the wide variety of scales and kinds of layering. Within these layers the
gradation of size, nature of clasts, and degree of sorting varies widely. These
textures are best preserved in kimberlites where the crater facies are preserved but
are also evident in deep parts of narrow pipes. Some of these varieties can be
ascribed to primary pyroclastic processes and some to redistribution of primary
constituents from reworking by normal sedimentary agents such as water, wind and
gravity. Terms such as pyroclastic or re-sedimented kimberlite may then become
applicable if the origin is not in doubt (Sparks, 2006).

Marginal wall rock breccias are the third prominent rock type. They range from
breccias composed entirely of wall rock to breccias with variable amounts of
kimberlite matrix of the MVK type. The latter vary from net veining of in situ wall
rock to clast-supported breccias to matrix-supported varieties.

The fourth type, hypabyssal kimberlites (HK), are those that have not been involved
in an eruptive process, and hence appear to be uniform in texture and coherent
(Skinner and Marsh, 2003). The country rock lithic clasts are usually strongly
altered due to the reaction between the fresh kimberlite magma and the rock into
which it intrudes. These rocks can be found in all parts of the pipe and are
characterised by abundant macro crysts and groundmass containing typical
kimberlitic minerals (E.g. monticellite, spinel, perovskite, calcite, and serpentine).

Understanding the makeup of each of these types of rocks is important both from a
sampling perspective, and to understand relationships between rock type and
physical properties.

50 Stephen Coward - April 2020

 Chapter 2: Literature review

2.2 Rock property sampling and characteristic measurements

Estimates of the physical characteristics of rocks within a kimberlite pipe are often
based on samples of the rock that are many orders of magnitude smaller than the
pipe itself. There are several processes used to obtain samples of the kimberlite,
including coring, chip sampling and bulk samples acquired from trenches or shafts.
The perceived cost of sample acquisition, especially in the case of early in the life of
projects where capital is rationed, is one of the main constraints on sample size,
sample count and the spatial distribution of samples within the kimberlite body.
These constrains often compromise or limit the representativity of the samples
acquired. Technical challenges of acquiring the rock mass required for testing have
also been documented (Hoek, 1997). The challenge is proportional to the variability
and scale of the orebody, and because kimberlite orebodies are complex,
proportional and representative sampling is challenging.

The objective of taking samples is to measure a portion of the total population and
acquire data that can be used to generate unbiased estimates of identified
characteristics of values at un-sampled locations. Sampling optimisation is the
process of identifying the set of tools and processes that will enable one to estimate
with least bias, imprecision, uncertainty and at lowest cost, the values, and
properties at un-sampled locations. This suggests that any value-optimising
business will try to balance the cost of sampling with the value of the information
obtained. The difficulty has always been to value the potential additional
information that will be acquired by the sampling before the sampling takes place.
Sampling technologies and practices reviewed are focused on the method of
sampling and measuring the characteristics of the in situ kimberlite that impact on
process performance in such a way as to facilitate the spatial estimation of the
characteristics that have been measured on the samples.

Metallurgical test work, as described by Lewis (2001), requires representative

samples from each defined lithology, or geological zone, which are prepared and put
through a laboratory scale test that resembles some aspects of operational scale
processes. The aim of these tests is to measure the rock's response to the applied
process (such as the bond work index (Bond, 1943)). The data acquire in this way
requires upscaling to predict operational scale performance. The scaling factor, or
factors, is commonly derived from experience of historical scale up that has been
required for other rock types or mine locations. The aim of this research was to
develop techniques to measure and/or sample the primary properties of the rock,
and its constituents, to be able to predict the relationship between the primary rock
properties and the expected process responses.

Stephen Coward - April 2020 51

 Evaluation of metallurgical recovery factors for diamonds recovered from kimberlites

This research has found, for kimberlites, that it is possible to sample the primary
properties of the rock, estimate and simulate these at block scale and then to use
these estimates to predict the operational scale process response.

Geostatistical techniques, such as ordinary kriging, can be used to generate

estimates for variables at unsampled locations (Armstrong, 1998). It is imperative
that sources of variation or error in the sampling as well as any bias is minimised.
Lower error in the input data would imply that the spatial models built on this data
would also have lower error. It is also possible to improve spatially estimated
models with integrated data acquired from different measurement techniques
and/or tools (Dowd, 1997).

It is important to understand and quantify the impacts that the nature of the variable
measured has on selection and execution of geostatistical estimation methods. The
'strange' behaviour of several response variables suggest that it is important to
validate assumptions that underpin the geostatistical estimation or simulation
method used. This includes, for instance being able to validate the relationship
between change of support and variable additivity (Carrasco et Al., 2008).The
implications here are that the design of sampling and testing programs for physical
rock characteristics should include collection of information that will allow the
testing of assumptions that are important to identify and select a validate the
estimation method. This is expanded on in section 2.3.

This research presents an experiment in which a combination of physical tests were

carried out with a description of each sample’s morphology, mineralogical content
and a number of non-destructive geophysical measurements. This sampling
framework has provided a data set that can be used to explore the some of the
challenges associated with the measurement and estimation/simulation of these
variables. Many of the rock characteristic variables required in the value chain
simulation exhibit characteristics that may hinder the generation of valid block scale
spatial models. These include:

• Scale - the scale of the sample and the scale of the estimate may be very
different. Although the sample data are considered to be "Point" values they
are in reality some statistic (e.g. an average) of a phenomenon collected or
measured over some support. This support (the size shape and orientation
of the sample) and the way the test integrates the measure over the support
will influence the characteristics of the data produced (means and
variances). When making predictions, either at sample support scale or far
larger scales the combination of support attributes of the sample and
estimated volume needs to be carefully considered. Change of support
calculations require a model of the relationships between distributions at

52 Stephen Coward - April 2020

 Chapter 2: Literature review

different supports. Deriving these models for complex distributions, non-

additive variables and variables that are ratios can be problematic;
• Additivity - if the variables are not additive then linear weighted
combinations of sample values will not produce unbiased estimates,
• Skew distributions - If the distribution of a variable has a high coefficient of
variation, is dispersed, and is skewed it is likely that a large number of
samples will be required to characterise the variable. Also, for this type of
distribution not only will the variance reduce as we take more and or larger
samples, but the mean of the sampled variable will shift. This has
consequences for change of support calculations;
• Non-linear relationships - If the relationship between the rock
characteristics and the measured response is non-linear, averages of the
responses will produce biased estimates. in some cases if the range of
variables is small the bias resulting from non-linearity may not have a
material effect;
• Short Ranges - if the variable at a given support has a short range, close
spaced sampling will be required. If this is not possible then the spatial
structure of covariance will not be able to be modelled, and will mean that
limited spatial estimation methods can be considered with resulting low
confidence in the spatial model; and
• Large Scale Phenomena - is certain cases the specified sample support for
a test may not be able to capture a phenomena that operates at a larger scale
e.g.., trying to measure the abundance of large clasts in a rock type where the
sample dimensions are close to the size of the large clast, or 2m fractures in
1m rocks.

These challenges reinforce the notion that where possible the preferred approach
is to spatially estimate or simulate so-called "Primary" additive variables wherever
possible (Coward, 2008). Many rock "response" variables are not additive, this
suggests that the variability and scale relationships are not simple, and hence their
spatial estimation and simulation will require rigorous demonstration that these
variables do indeed meet with the requisite assumptions of the estimation and
simulation techniques used.

Implications of the Theory of Sampling

The development of the Theory of Sampling (TOS) in a mineral context is attributed
to Pierre Gy, whose first paper on the subject was published in 1950. Although most
of the test work presented here was carried out on in situ materials that are not
necessarily ‘broken ore,’ the concept of discrete materials relates clearly to scale.
“All matter is, at an atomic level, discrete and continuity is in essence a
mathematical concept" (Gy,2004). This is of critical importance when making
measurements of rock properties on kimberlite cores which may only be

Stephen Coward - April 2020 53

 Evaluation of metallurgical recovery factors for diamonds recovered from kimberlites

representative of one of the many components that make up kimberlite (e.g., a

country rock clast that contains no kimberlite matrix material) rather than of the
'mixture' of components that will drive recovery. In extreme cases, it is easy to
identify flaws in sampling protocols, but it becomes more difficult when the
heterogeneous nature of kimberlite creates subtle biases and errors.

A taxonomy of sources of error is central to the concepts that underpin the theory
of sampling. These are then used to determine a way to relate the mass of the sample
to the expected error in the estimate. The Discrete Selection Model formalised and
published in 1975 has recently been updated (Gy, 2004) and the relationship
between sources of error can now be depicted as in Figure 10.

Global Estimation Error (GEE)

aR − aL
GEE =
aL

Total Sampling Error (TSE) Total Analytical Error (TAE)

a − aL aR − aS
TSE = S TAE =
aL aL

True well defined proportion of component

aL A in lot L, always unknown
Primary Secondary aR
Analytical result, an estimate of a S and
hence estimator for a L
Sampling error Sampling error
(PSE) (SSE) aS True well definable proportion of component
A in sample S, always unknown, also and
Occurs in removal from Occurs in laboratory estimator of a L
production in plant

Correct Sampling Errors (CSE) Incorrect Sampling Errors (ISE)

Mass reduction with uniform Mass reduction with non-uniform
probability- but material heterogeneity probability – incorrect equipment
drives variance

Fundamental Grouping and Incorrect Incorrect Incorrect

Sampling Error Segregation delimitation Extraction Preparation
(FSE) Error (GSE) (IDE) (IEE) (IPE)
Error from the Error from Geometrical Extraction of Preparation and
constitutional distributional delimitation of 3d Geometrical of 3d handling of
heterogeneity heterogeneity increments increments material
incorrectly incorrectly increments and
sample

GEE= TSE +TAE =(CSE+ISE) + TAE = (FSE+GSE)+(IDE+IEE+IPE)+TAE

Figure 10: A hierarchical depiction of the relationship of sources of error arising from
sampling, Adapted after Gy (2004).

The dominant error is taken to be the fundamental error, although the others should
be minimised as far as possible. The primary cost driver in processing samples is
likely to be size, thus using this approach to sample design aims to minimise the size
without increasing the fundamental error to such an extent that the information
obtained is of negligible use in estimation. To this end Gy proposed the formula for
deriving the Fundamental Sampling Error. A simplified version is presented in
Equation :

54 Stephen Coward - April 2020

 Chapter 2: Literature review

C d3
 ( FSE ) =
2
Equation 3
Ms

Where

C is the sampling constant, consisting of four sub-constants,

d is the top particle size (upper 95% average grain diameter)

𝑀𝑠 is the sample mass

Gy proposed several simplifications of this calculation, with associated assumptions

concerning the constants that might apply. Gy’s work identifies errors associated
with particle size as being the most important. Most of the experimental sampling
in this research has been on in situ material. At some clast size the heterogeneous
nature of the material being tested will dominate the variance of the results
obtained. One potential approach is to determine the end member response of the
country rock clasts and that for pure kimberlite. It would then be possible to define
the maximum acceptable size of clast for a specific variance in the measurement
taken. This methodology is a version of the method suggested by Royle (1986).

Errors will arise from many sources; hence effort should be focussed on identifying
and eliminating sources of error when sampling. Where the errors cannot be
eliminated, the estimate of the magnitude of the relative errors must be considered
in any consequent data analysis that is carried out.

The properties selected and identified in section 1.4 include those that can be used
to model and predict kimberlite comminution, diamond liberation and effectiveness
of separation processes. They can be broadly described as either primary or
response variables as defined by Coward (2009).

Early project stage sampling

From early stage exploration through to the beginning of production, the dominant
source of material characterisation data is from downhole logging, core and chip
samples that are subjected to several non-destructive and destructive tests.

Work carried out by Copur and Billgin et al. ( 2003) on a range of rocks yielded the
data that is depicted in the plot in Figure 11. Although the range of rocks tested are
not kimberlites the work demonstrates how it is possible to collect different types
of data to explore the existence of a relationship between the primary variables (e.g.,

Stephen Coward - April 2020 55

 Evaluation of metallurgical recovery factors for diamonds recovered from kimberlites

various textural measures, measures of rock content) and response variables . The
nature of these relationships is clear for some rock types and not for others. This
suggests that for some rock types the variance of rock responses within the 'rock
type domain' may well be greater than the variance of rock characteristics between
'rock type domains'.

Associations between properties for several rock types

Unconfied
(MPa)

200 14 Compresive strenght

UCS
180 Static Elasticity
12 Modulus Esta
Measured values

160
140 10 Dynamic Elasticity
modulus Edyn
120 8 Schmidt hammer
100 rebound value Shrv

80 6
Density Y
60 4
40 Brazillian Tensile
strength BTS
2
20
Acoustic P wave
0 0 velocity Vp
0 1 2 3 4 5 6 7 8 9 10 11 Cerchar abrasivity
value CAI
Rock types

Figure 11: Results of a number of rock tests on a range of rock types after Copur et al.
(2003).

The use of downhole measures and laboratory scale tests relevant to this research,
such as acoustic velocity is described more fully in section 4.2. The outcomes of
early stage sampling are used in desktop studies to validate additional exploration
of the target.

Feasibility and production stage sampling.

To create a sufficiently robust business case to justify the funds required design and
build diamond mines requires reduction of the uncertainty in the value of diamonds
in the intended recovery size envelope. as these often have a large impact on the
mine's revenue. To acquire a parcel of diamonds that is representative of the
diamond size distribution, especially diamonds > +10ct size fraction, often requires
bulk samples of the order of several thousands of tonnes. During the treatment of
these bulk samples, several process parameters and performance characteristics
are recorded. This data is gathered to describe the kimberlite's impact on five areas
of the diamond recovery process including:

1. Comminution, with measures of rock hardness/competence and amenability

to various forms of comminution;
2. Dense media separation, with measures of dense particle abundance in
specific size ranges;

56 Stephen Coward - April 2020

 Chapter 2: Literature review

3. Fines, ultra-fines and their impact on water handling, with measurements of

modal mineralogy in appropriate size fractions, especially of smectite clay
minerals, and extractable sodium percentages in process water-clay slurries;
4. X-ray luminescent intensity of gangue minerals in DMS concentrates with the
potential to increase diamond-by-weight percentages in the final recovered
product; and
5. Magnetic properties of concentrates from DMS that have the potential to be
stripped from concentrates before final recovery thus reducing quantities of
material being processed through final recovery.

Each of these aspects may include qualitative descriptions of the characteristics of

the kimberlite fed to each process, the sampled size distribution and measures of
the rock characteristics such as density where appropriate.

These data can be used for process design, informs the geological model and
improves the models of the relationships between primary and response variables.
The data is however usually restricted in spatial coverage as the bulk samples are
likely to be selected from single accessible locations in each major kimberlite
domain

2.3 Challenges of spatially estimating physical rock

characteristics

Data obtained from measuring the physical characteristics of samples can be used
to estimate the characteristics of the un-sampled portions of the orebody. These
variables, unlike grades, may exhibit various behaviours that make their estimation
at block scale somewhat challenging.

There are several ways to generate spatial models of variables at different scales,
the methods considered here include:

• Averages of the samples

• Polygonal estimates
• Geostatistical estimation techniques such as ordinary Kriging
• Spatial Simulation using geostatistical techniques

Using sample averages to estimate rock properties

It is possible to process the data rapidly by using the average of relevant sample
values as the estimate for the variable of interest. This however does not consider
the spatial correlation of either the samples or the locations and volumes that are to

Stephen Coward - April 2020 57

 Evaluation of metallurgical recovery factors for diamonds recovered from kimberlites

be estimated. However, if the rock is very homogenous and the sampling has been
carried out in a very regular grid, then this may be an appropriate approach.

Although this approach is computationally efficient it does not consider the spatial
location of samples, does not account for the differences of the scale of the sample
vs the scale of the estimate, and does not provide a means to assess the quality of
the estimate made.

Polygonal estimates
This approach utilises the generation of a volume that the sample is deemed to
represent. The region of interest is divided into shapes that fill the volume, in a way
that ensures that each volume has at least one sample in it.

This approach may be appropriate when there are sufficient samples regularly
located throughout the region of interest. It is computationally efficient and does
not require any modelling of functions.

Kriging
Kriging is a term used for a family of estimators whose origins were attributed to
Danie Krige by George Matheron (1963). Matheron (1963) named the process used
to generate estimates "Kriging" as a tribute to the work of Danie Krige. Ordinary
kriging is a linear estimation algorithm for a regionalised variable that satisfies the
intrinsic hypothesis.

The benefits of this approach are that the estimate of the variable of interest at
unsampled locations is unbiased, and that the error of the estimate at sampled
locations is calculated during the estimation procedure.

Spatial simulation
The evaluation of the metallurgical recovery factor requires an understanding of
both the average, expected, and variable nature of the rock characteristics on the
recovery factor. One approach to overcome some of the limitations of block scale
estimates is to generate a sufficient number of equally plausible realisations of the
orebody characteristics and then to use these as inputs to a process simulation. The
set of simulations represents the uncertainty in the ore characteristics and each
individual realisation can be used to explore the impact of block to block variability
on process performance.

58 Stephen Coward - April 2020

 Chapter 2: Literature review

These rich spatial models can be created using one of several spatial simulation
techniques (Dowd, 1994). Spatial simulations aim to reproduce the histogram of the
data and the spatial co-variance of the data.

Selection of a combination of methodologies is required to develop spatial models

of several characteristics of the kimberlite deposits that can be used in the
simulation of the operation of kimberlite mines. The process flow that this research
has followed is similar to that used by (Dowd and Dare-Bryan, 2004) and is depicted
in Figure 12. Both estimated and simulated orebody models can be generated from
the original real sample data.

One of the realisations of the simulation set (based on real data) is selected at
random and deemed to be the "Real deposit. This model can be sampled using
different sample grids to generate data that represents the results of sample
experiments that have different geometric layouts and sizes.

These 'virtual samples' from the 'virtual orebody' can be used to generate new
estimated orebody models and condition a new set spatial simulations of the
kimberlite deposit.

Real Sample
Estimated
Data
Orebody Model
Mine and
Process
model
Simulated
Estimated Simulated Orebody
Orebody Orebody
Model Model

Estimated
Select One Orebody Model
Realisation: Mine and
The ‘Virtual Process
Orebody’ Model
Simulated
Orebody

Extract Several Estimated

Sample Sets Orebody Model
Mine and
Process
Model
Simulated
Orebody

Figure 12: Schematic of the process used to create a framework to quantify the impact of
changes to sampling , estimating and simulating Kimberlites on process models.

Stephen Coward - April 2020 59

 Evaluation of metallurgical recovery factors for diamonds recovered from kimberlites

The ‘Virtual orebody,’ (i.e. the one selected realisation taken from a simulation set
based on real sampling data ) can be taken to represent so-called 'perfect knowledge
case' against which various comparisons can be made (e.g., the impact of sample
spacing, simulation method used, crushing process model etc.). The method has a
few limitations and is computationally demanding but allows for a myriad of
possible comparisons. The framework or platform is used to compare several
aspects of the methods and techniques presented in this research, including a
comparison between estimates of expected recovery based on limited information
with that which would have been realised had the ‘virtual orebody’ been treated.

The techniques that can used to generate spatial simulations include turning bands,
sequential indicator and sequential gaussian simulation. Each has their own benefits
and drawbacks; a brief description 'turning bands' which has been used in this
research is described.

Turning bands simulation was developed in the early 1970’s (Dowd, 2004). This
method of simulation uses a one-dimensional random function that is set up along
lines that have an even spatial density. The simulations from these lines are then
projected out onto a grid that is set up around the lines. Some of the limitations of
this approach include the generation of artefacts if insufficient lines are used. The
method deals with anisotropy indirectly, usually through grid expansion and
shrinking, and the method requires a separate kriging step to condition the
simulation.

Implications
The method used to generate spatial estimates and simulations of rock
characteristics has consequences for the use of these models in value chain
simulation. Estimates derived using linear methods such as ordinary kriging are
likely to underestimate the variability of orebody variables such as grade that will
be observed during operation. Simulated realisations of variables of interest at block
scale will better reflect the variability of the characteristics of the blocks as they are
processed. It is however important to consider that the scale of measuring rock
characteristics (typically the size of core) is many orders of magnitude smaller than
the size at which the process will respond to the change in the specific
characteristics.

The approach described in this thesis suggests that matching the support of the
sample, the location and spread of samples in the pipe to be evaluated and the scale
of the spatial estimate is a very important aspect that will determine the effective
use of this important input into value chain modelling.

60 Stephen Coward - April 2020

 Chapter 2: Literature review

2.4 Mineral process models and simulation practices

Mineral processing, such as crushing and dense media separation, can be modelled
using several mathematical techniques. King (2000) provides a demonstrated
taxonomy of systems together with a classification of these with benefits and short
comings. The emerging use of so-called “expert systems” integrated with dynamic
simulation for use for a) calibrating process simulation models and b) for improving
operational decision making.

The processes used to recover diamonds from kimberlite can be divided into mining
and treatment processes. There are several levels of process modelling and
simulation, from 'black box simulations', that merely transform an input to an
output through a simplistic mathematical function, to complex systems, that address
the interaction of individual particles and elements in very small-time increments.

The objective in this research is to identify current practices in mining and

treatment simulation that can be validly used in a value chain model to analyse the
metallurgical recovery factor. The models that have been considered for use in this
value chain modelling need to meet with the following criteria:

• The models respond appropriately to estimated and simulated rock

characteristics; and
• The models can be used in a system simulation that can be run in a realistic
time frame.

The primary objective of the integrated evaluation model is to create a value chain,
comprising linked models, that can be used to simulate the operation of the mine
and treatment plant system. The mining model translates the estimate of in situ rock
characteristics into a sequential ore stream. The ore stream characteristics are used
to drive process models which are used to calculate the plant throughput and
efficiency and derive an estimate for diamond production based on the
characteristics of the incoming ore stream.

Models of mining processes

Kimberlite pipes are mined using open pit and underground methods.

The design, or planning, of an open pit mine has three primary components:

1. Determination of the geometry of the ultimate pit shell;

2. The feasible extraction sequence of the blocks; and
3. The scheduling of block extraction for mining.

Stephen Coward - April 2020 61

 Evaluation of metallurgical recovery factors for diamonds recovered from kimberlites

Several algorithms have been developed to optimise the process of mine design,
including the Lersch Grossman approach developed from graph theory (Lerchs and
Grossman, 1965). The aim of the optimiser is to achieve the maximum present value
by offsetting estimates of cost to mine the block against the value recovered from
the block. The estimate of block value is usually derived by multiplying the
contained mineral in the block by the price that will be obtained for the mineral
minus the cost of extracting, processing and delivering the valuable content of the
block. The associated present value can be obtained by developing and applying a
feasible production schedule.

Underground methods used to mine kimberlites include block caving and room and
pillar methods. These methods are subject to many geometric and ground condition
constraints. The processes used to derive the geometry and sequence of the
operation are beyond the scope of this research.

For the models considered here it is assumed that an optimal mine plan has been
designed for the simulated deposit based on the kriged estimate of recovered value.
Re-running the mine design for each realisation would not only be time consuming
but is not appropriate given that the simulations are designed to maximise the
variance at block scale. There is however ongoing work to incorporate uncertainty
in the derivation and optimisation of the ultimate pit shell, block sequence and
schedule (Dowd and Dare-Bryan, 2004, Lamghari and Dimitrakopoulos, 2016.).

Mining is the first stage of the process where ore breakage occurs. The processes
used include a combination of primary drilling and blasting. An approach aimed at
optimising selectivity during the drill and blast cycle has particular relevance to this
research (Dowd and Dare-Bryan, 2004). The method is centred on the creation of
an in situ model that represents “perfect knowledge at all scales” through
simulation. This simulation is then sampled on a user-defined grid to provide the
limited information that would normally be available for planning. This information
can then be used to design the blast and muck pile clearing pattern. It is then
possible to evaluate how well the plan is likely to perform by comparing what was
selected and trammed to the plant given the mixing caused by blasting, against the
expectation derived from simply using the estimated block characteristics of the
planned mining blocks.

Although the approach has been designed to optimise selectivity, the simulation
approach used to create estimates at relatively small scales, far smaller than the
smallest selective mining unit (SMU), demonstrates that rock properties can be
simulated into blocks of an appropriate scale. The key limitation is the minimum
size of the block that can be simulated in the blast process. The minimum
dimensions are driven by the resolution of sampling, and the processing power to
simulate this process.

62 Stephen Coward - April 2020

 Chapter 2: Literature review

Methods aimed at optimising the size distribution of material fed to a mineral

processing plant have been developed into a multi-organisational project ‘Mine to
Mill’ and several of the developments in blast modelling have relevance to this
research (Napier-Munn et al., 1999). Central to this approach is the balancing of the
cost of fragmentation achieved through blasting and the cost of fragmentation in the
process plant. The study has led to the development of non-contact visual
technologies that can be used to monitor fragmentation in the mine and in the
processing plant.

Mineral processing modelling

A research survey carried out across the De Beers group of South African mines,
obtained from senior process managers their views of which ore properties would
influence process design, cost and efficiency, identified the following variables:

• Brittleness, strength and plasticity – impacts crushability and resulting ore

size distribution hence mass balance, other derived measures such as nip
angle were also mentioned.
• Clay type and content – impact on mass balance, water consumption and
treatment requirements, also impacts on the Dense Media Separation (DMS)
process.
• Ore Density distribution – the relationship between changes in size
distribution and the change in density distribution of the ore stream impact
both mass balance and yield from the DMS process. This also affects the
efficiency of DMS when the enclosed minerals have a narrow size
distribution.
• Nature of dense species, which will sink in the DMS, in the ore and their
impact on recovery processes –includes luminescence distribution, magnetic
susceptibility, and porosity.
• Diamond characteristics – including the size distribution shapes and
luminescence distribution.

Process Function Material Property Impact on Process

Bulk Fabric, micro fractures, Faults Determines blasting characteristics
Hardness Change in power consumption, and size distribution
Liberation Brittleness Change in size distribution, better fragmentation
Clay content Higher clay requires more water, slower settling rates
Friability Dust creation, scrub ability
Density distribution Impacts on separation, and yield
Separation Shape of rock particles Impacts on effective cut point
Size distribution Screening efficiencies and separation efficiency
Gangue Magnetic Susceptibility Impacts of magnetic separator yield
Gangue X-Ray Luminescence Impacts on X-Ray through put and efficiency
Recovery
Diamond Magnetic Susceptibility Impacts on Magnetic separator efficiency
Diamond X-Ray Luminescence Impacts on X-Ray through put and efficiency
Table 4: Relationships between ore property and the impact on mineral processing.

Stephen Coward - April 2020 63

 Evaluation of metallurgical recovery factors for diamonds recovered from kimberlites

In a review of current plant design practice, a listing of typical plant specification

variables was obtained (Lilford, 2003). Several of these specifications are based on
data obtained from samples taken from the orebody to be treated. Although there
are numerous input variables( e.g., rock hardness, UCS etc.), the design criteria are
often based on the maximum, average and minimum values of these variables. These
are obtained from a set of samples deemed to be 'representative' of the orebody. It
is suggested that the relationship between the sample values, their size and the
design criteria affected is an area for further investigation.

In 2000 an analysis of operational cost databases indicated that the highest

individual costs on average for the suite of De Beers operations were; manpower,
electrical power, and ferrosilicon (Shipiki-Kali, 2000). Most of the electrical power
was consumed by comminution devices. Subsequent analysis of the cost data for
2003 indicates that the major cost drivers have not changed. This finding implies
that the data obtained from sampling for kimberlite characteristics may also be
useful for improving the estimates of power consumed for comminution and in
reducing the uncertainty in the prediction of operational cost.

Modelling comminution
The modelling of comminution has received much attention by several authors
(Agricola, 1556; Bond, 1943; Witten, 1972; Napier-Munn et al., 1999). To model
comminution requires a generalised formulation for the mechanisms that drive the
reduction in the size of rocks. When rocks are crushed the mechanisms of size
reduction include:

• Attrition;
• Chipping;
• Impact fracture of rock particles; and
• Breakage of contained mineral particles.

King (2001) proposed a generalised model based on assigning weights to several

probabilistic breakage models. The weights and breakage models are calibrated to
suit both the material being crushed and the device that is used. An example of a
breakage model for a mill is given below to show the composition of the breakage
functions. Essentially this is comprised of the summation of the proportion of
material arriving and leaving each size fraction per unit time across all size fractions.

Specific attrition rates are ore specific. Techniques developed by the Julius
Kruttschnitt Mineral Research Institute show that this can be determined by
treating a sample for 10 minutes in a tumble mill and the progeny screened (Napier-
Munn and Morell et al., 1999).

64 Stephen Coward - April 2020

 Chapter 2: Literature review

An attrition breakage function can be obtained from assessing the slope of the graph
depicting the cumulative size distribution of the finer fractions. The change in mass
(m) in a size fraction per unit time is expressed in Equation 4 :

dm m Equation 4
= −3k
dt d f− 
1

Where

dm is the change in mass per unit time;

𝑚 is the mass in a size fraction;

𝑑𝑓1−∆ 𝑖s the geometric mean of the feed size distribution; and

k is a material-specific constant

Integration of this equation with respect to time will determine the total
comminution, or size reduction that will be achieved for a given unit process.
Comminution modelling and its use in the value chain model is described more fully
in section 6.3

Modelling dense media separation

This process uses the difference in density between kimberlite and diamond to
separate the diamonds from gangue material. If the overall size distribution
achieved by the comminution section can be calculated, then this ‘Total Grind’ can
be used to estimate diamond liberation and recovery. The ‘Diamond Liberation
Granulometry Model,’ estimates the lock up in different diamond classes and
predicts what proportion of the recovered diamond distribution is lost in the
tailings from the DMS circuit. These losses can be used to estimate the metallurgical
recovery factor. The model can be used to calculate the proportion of each diamond
size class that will be liberated given an estimate of the total grind size distribution
of the kimberlite. As the model works across all diamond size classes it is also
possible to use the model to assess the impact of changed liberation on the
recovered diamond size distribution.

The model was developed by Kleingeld (1976) in an attempt to estimate the locked
diamond content of cemented gravels that were discarded by coastal operations in
southern Namibia.

Stephen Coward - April 2020 65

 Evaluation of metallurgical recovery factors for diamonds recovered from kimberlites

The purpose of the model is to produce a reliable estimate of the diamonds that
might plausibly be contained in the discarded rock stream based on the operating
parameters of the plant and the size distribution of the recovered diamonds. Each
size class of ore particles (for example the -10mm +8mm size fraction) has a
potential to float out of the dense media during the separation process and report
to waste. The particle class that is discarded can carry with it a size of diamond that
can be determined through the density relationship between diamond, the rock that
it is enclosed in, and the effective density cut point at which the separation is made.

The model requires the following inputs:

• the total grind achieved estimated from ore stream size distributions;
• the number of ore particles in each size fraction;
• the average ore density in each size distribution; and
• the recovered diamond distribution.

The probable locked diamond distribution is calculated by assessing the proportion

of each particle class volume that could potentially be diamond, so that the
combined particle density is equal to the medium density. This locked distribution
can be added to the actual distribution to create a predicted total plant feed diamond
distribution. The difference between the two gives the possible additional diamonds
that could be recovered assuming that 100% liberation and recovery is achieved.

The model is applied during the sample phase to estimate the liberation during bulk
sampling. Although the estimate of locked diamonds is not added to the modelling
phase of the total content distribution, it is used as a qualitative guide to the
difference in liberation that is likely to be achieved between the sampled grade and
the expected production facility.

Currently it is not common practice to determine how the range of rock properties
will vary in the short term and how this will impact on the efficiency and throughput
of the process. In most cases, the global average rock properties are used to estimate
the expected ore grind size distribution and diamond liberation.

The research reported in this thesis seeks to demonstrate that that the ‘Diamond
Liberation Granulometry Model,’ can be adapted and used effectively to integrate
total diamond content models with spatial models of kimberlite properties to
evaluate diamond recovery.

66 Stephen Coward - April 2020

 Chapter 2: Literature review

2.5 Approaches to evaluating uncertainty and variability in

diamond recovery

Published accounts of estimated diamond recovery factors often only report a single
value for expected grade recovery (Farrow, 2015). A single point estimate of
expected, or most likely, recovery of the contained diamond content does not give
any indication of how uncertain or variable the diamond production will be over the
life of the operation.

One important source of uncertainty arises from the design of sampling campaigns
to define reserves. These campaigns are often constrained by the capital allocated
to sampling the orebody. The optimal selection of the size, number and geometric
location of samples to achieve a specified degree of confidence in the orebody grade
cannot be determined with any degree of accuracy prior to executing the sampling
campaign. There are some generic 'rules of thumb' that can be applied to the
classification of resources and reserves (Parker, 2000) based on the expected
variation in production that can be inferred from the quality of a completed resource
estimate. The metrics used in classification range from search neighbourhood size,
drill hole spacing, and/or kriging variance (Silva and Boisvert, 2014). It is however
difficult to predict these prior to executing the sampling campaign and performing
the estimate. Some inference can be made from early geological modelling or from
other kimberlites with similar architectures to predict the resulting uncertainty in
estimates of in situ volume, density, tonnage and grade, but it is far more difficult to
predict the recovered values for these variables.

An alternative approach to sampling campaign design would begin with defining the
tolerable financial uncertainty for a project and then use these limits to define the
required degree of confidence in the variation in expected recoverable reserves. To
do this however requires estimates for several technical factors and a methodology
that can be used to calculate how the in situ values, which are spatial in nature, will
be transformed into revenues over the life of the project.

The understanding of tolerable financial uncertainty is determined, to an extent, by

a combination of listing code requirements such as JORC (2012) and SAMREC
(2016), and the availability and structure of the finance that will be used to fund the
project. The variance of the cash flows will be driven by the variance in the estimate
of reserves, the ability to mine and extract the mineral commodity and the range of
the prices that the mineral could realise. For illustrative purposes, research
reported in this thesis has assumed that the commodity price and ability to market
and sell the diamonds is given. In the remaining two areas that influence financial
variability, the research evaluates several methods to model the uncertainty in the
recovery factor and its variability. The approaches are described in chapter 6 and

Stephen Coward - April 2020 67

 Evaluation of metallurgical recovery factors for diamonds recovered from kimberlites

demonstrated in chapters 7 and 8. This work shows how it is also possible to derive
confidence limits for the project value based on predicted variability in project
cashflow.

Calculating the expected return of a mineral deposit is a multistep iterative process.

The output of this process is used to guide investment decisions. There are many
valuation techniques that can be used to create a financial model for mineral
deposits, but all require an estimate of future cash flows (Table 5)

Development Preferred methodology Approach

stage

Mineral Rights Market comparables or nominal rate Cost of Sales

comparables

Early exploration Multiple of expenditure, R/Ha or $/oz market Cost of sales

comparables comparables

Advanced Discounted cash flow, option pricing Income option

exploration pricing

Pre-feasibility Discounted cash flow Income

studies

Feasibility studies Discounted cash flow Income

Development Discounted cash flow Income

stage

Production stage Discounted cash flow Income

Salvage or closure Appraised value or historical cost Cost of sales

comparables
Table 5: A list of preferred valuation methodologies used in South Africa (Silva and
Minnitt, 2005) .

There are, however, moves within the industry to attempt to create guidelines for
this process. This includes the international valuation standards committee as well
as several other bodies around the world. Although it is a guideline and not
prescriptive SAMVAL (2016) suggests:

“Proven and Probable Reserves valuation should be based on the application of

appropriate, disclosed modifying factors, as applied in the Life of Mine plans of the
operations or company, with historic cost reporting as an alternative.”

Although not prescriptive on the methods used it is suggested that the three bases
for valuation should be income, sales comparison and cost. A brief description of the
assumptions and limitations of methods used to value mining projects is presented

68 Stephen Coward - April 2020

 Chapter 2: Literature review

below. The review also demonstrates how to incorporate the metallurgical recovery
factor into financial models and hence demonstrate the financial impact of the range
and uncertainty of the recovery factor in financial terms.

Times pay back and duration of payback methods

One of the simplest methods of evaluating a mining investment is to calculate the
number of times that a project will pay back the initial investment and the time that
this is expected to take. The method is particularly useful at an early stage of
evaluation when there are many scenarios to compare. The method can be
expanded by discounting the cash flows at a certain interest rate to account for the
relative values of cash flows received at different times.

Discounted cash flow methods

The principle of discounted cash flow assumes that cash flow today is worth more
than the same cash flow tomorrow. The change in value is a function of both
potential depreciation in value as well as a measure of uncertainty of the cash flow
tomorrow. This necessitates the use of a discount rate. The selection of an
appropriate discount rate is key to all discounted cash flow methodologies. The rate
is sometimes referred to as the hurdle rate, or the opportunity cost of capital. The
discounted value of a cash flow at some time in the future can be expressed as in
Equation 5.

DCF =
 cashflows i
Equation 5
(1 + r ) i

Where

DCF is the discounted cash flow

r is the discount rate

i is the period number

The model described above can, and often is, used in deterministic fashion to assess
the relationship between an input variable and the value of the project; This is
known as sensitivity analysis. Sensitivity analysis may be a suitable technique for
assessing relationships between input variables and output values for relatively
simple valuation models.

Stephen Coward - April 2020 69

 Evaluation of metallurgical recovery factors for diamonds recovered from kimberlites

By flexing one parameter at a time (termed 'independent sensitivity analysis') it is

possible to calculate project value for a range of values for the input variables, for
example it would be possible to show how the forecast discount rate for a project
with a lot of debt has a substantial impact on the calculated project value. This is
repeated for several of the input variables. The analysis then ranks the importance
of the variables based on the 'sensitivity' of project value to proportional changes in
the values of the input variables. The assumption is made that the variable that is
most important (or presents the most risk to the project) is the one that results in
the biggest change in project value for the smallest change in input variable,

As valuation models include more variables to improve the 'reality' of the model,
independent sensitivity analysis is increasingly likely to incorrectly identify the
variables that that are going to have the largest impact on project value. To mitigate
this risk, it is possible to build in correlations between the input values e.g. as
interest rates rise so do wages and perform a 'correlated sensitivity analysis'.

The 'sensitivity centric' approach to assessing project risk often gives a better
indication of the structure of the model itself rather than providing a valid way of
quantifying the relationships that exist between the orebody uncertainty and
variability, project configuration and project outcomes. The main shortcoming in
this approach is that the ranking does not account for the probability distribution of
the range for each of the input variables.

A slightly more complex technique is to use input distributions, that may or may not
be correlated, to generate the range of input variables used in the project model. The
parameters used for the distributions are selected in a qualitative manner using
expert opinion (Vose, 2002, Davis, 1995) as discussed further in section 9. In this
approach the type of distributions for important variables, and the values of their
parameters, are elicited from a panel of experts that can be weighted in several
ways (Aspinall, 2018). The limitation of this approach includes the biases that arise
during the process of eliciting expert opinion and potential biases that arise from
underweighting of extreme events, and overweighting of recent events (Kahneman,
2011). The use and application of these approaches are reviewed in more detail in
chapter 9.

Sensitivity analysis is carried out by drawing values for the input variables from the
calibrated distributions, and rerunning the model, in the order of thousands of
times. The distributions of outcomes can then be analysed to provide a far richer
understanding of the variable relationships that exist in the project. This approach
is sometimes referred to as 'monte carlo analysis'.

There are cases when two or more variables oscillate simultaneously, or are
correlated in a complex relationship over time , for instance the relationships that

70 Stephen Coward - April 2020

 Chapter 2: Literature review

exist overtime between the grade of the material mined and the cost of processing
that material. Prediction of the financial performance of complex systems require
models of appropriate sophistication (Dowd, 2015). The 'integrated evaluation
model' techniques described in this thesis have been developed to address the
complexities in such cases. It has been demonstrated that for mining projects this
method is likely to outperform sensitivity analysis (Nicholas et al., 2007). A more
recent evolution of the approach termed "Scenario Based Project Evaluation"
(Coward, et al., 2013) incorporates a more sophisticated approach to deal with
future operating environments as suggested by Bradfield et al.(2005).

2.6 Summary of practices and limitations identified in this review

Kimberlites are formed through several high energy processes over a long
geological timeframe (Field, 2009). This gives rise to deposits with a variety of
geometries and the domains within the deposits vary greatly in shape and size. They
contain many diverse minerals and exhibit a wide range of physical characteristics.
Mineralisation and mineral assemblages mean that the variables of interest can
have very different spatial covariances and that these can change over very small
ranges.

This research describes how the variability and uncertainty of the estimates of
mineral resource contents and physical characteristics can be used to improve the
estimate of uncertainty in production output and express this uncertainty in
financial metrics.

Simulation outputs of the mining and treatment processes includes an estimate of

the tonnage that can be treated and number of diamonds and their size distribution
that will be recovered. This output is used to generate a cash flow estimate that can
be used to value the deposit in several ways (Dowd, 2002). Although this research
is not focussed on developing new financial evaluation methods it is important that
the complexity of the financial model and value indicators used have an appropriate
degree of sophistication. The approach demonstrated here makes it possible to
show how uncertainties in resource estimation and reserve evaluation can be
expressed as risk in quantified financial terms. The aim here is to select several
indicators of project performance, ranging from valuation of very simple production
outputs(e.g., value of diamonds produced in a period) to more complex financial
metrics (e.g., Internal Rate of Return) to demonstrate how the impact of implicit and
explicit treatment of the recovery factor in the valuation process can be assessed.

The literature review clearly shows that the recovery factor has a material bearing
on the value of diamond projects. Furthermore, technology now exists that will
enable sufficient cost-effective samples to be taken from orebodies to enable

Stephen Coward - April 2020 71

 Evaluation of metallurgical recovery factors for diamonds recovered from kimberlites

estimation and simulation of physical characteristics of the kimberlite that will drive
mining and diamond extraction process efficiency. It is possible with some
additional work to link these directly to mining and metallurgical process models
and to use these to determine the range and variability in the metallurgical recovery
factor. These data can be used in appropriately sophisticated financial models to
express the impact that the range of metallurgical recovery will have on mining
projects in financial terms.

The remainder of this thesis covers experimental data acquisition, mathematical

modelling and value chain simulation to demonstrate and validate this approach.

72 Stephen Coward - April 2020

 Chapter 3: Problem statement

3 PROBLEM STATEMENT

3.1 Introduction

The factors that influence the expected value of a kimberlite diamond mine include
the proportion of the total contained diamond population that will be recovered, the
diamond sales price and the cost of mining and recovering the diamonds. Diamond
recovery and loss is a function of the effectiveness of the mining and metallurgical
processes. This research addresses important shortcomings of current methods
used to derive and quantify the range and uncertainty of metallurgical recovery.

Incorrect estimation of the metallurgical recovery factor impacts on decisions made

throughout the life of diamond mines. It has potential to mislead investors and
project owners, result in design of inappropriate processes, implementation of poor
strategies and can ultimately result in failed projects.

The challenge of quantifying the range and uncertainty of diamond recovery impacts
on the evaluation of mining projects from early discovery phases through to
operating mines and into project closure. Errors in predicting the recovery of
diamonds arise from the complex process required to acquire kimberlite orebody
information, the low concentration of the mineralisation and the difficulty of
sampling and estimating the physical characteristics of the kimberlite that impact
on process rate, process efficiency and cost.

This research presents a framework to address the primary shortcomings of the

traditional approach to deriving and using the metallurgical recovery factor. The
framework facilitates quantitative assessment of strategies that aim to reduce
diamond mining project risks that arise from uncertainty in the metallurgical
recovery factor.

3.2 Current limitations in the derivation and uses of the

metallurgical diamond recovery factor.

Large kimberlitic diamond deposits are often formed by multiple phases of

intrusion and subsequent reworking. The resultant diamondiferous orebodies
exhibit complex geometries, variable diamond concentrations (Field, 2008) and a
wide range of physical rock characteristics.

Stephen Coward - April 2020 73

 Evaluation of metallurgical recovery factors for diamonds recovered from kimberlites

Due to the range of rock types contained in these orebodies, the granular nature of
the mineralisation and its very low concentration, sampling of the grade and size
distribution of in situ macro-diamonds (diamonds >0.5mm) is time consuming and
costly (Ferreira, 2013). The accuracy of grade sampling that is carried out is subject
to biases that can arise during sample acquisition and treatment.

Bulk sample plants crush and separate diamonds using similar techniques to full
size production scale diamond processing plants. The impact of the rock
characteristics on the variability of recovery efficiency at sample scale is different to
that at full production scale. The relationships at different scales exhibit non-linear
relationships that are difficult to model and adapt for scale.

The technology selection, process configuration and operating philosophy of

diamond process plants is often based on long-run averages of the expected feed
characteristics and expected diamond characteristics (Morell, 2003; Daniel, 2010).
In some cases, these designs are adapted to accommodate an ‘operating envelope’
of feed rock characteristics. There is however little indication in the literature
reviewed of the use of short-term variability to improve the design and operation of
processing plants.

Process simulation capability continues to develop and improve with advances in

computing processing capacity, processing speed, and several advances in
processing algorithms (King, 2001). Historically, however, they have been limited in
several ways by the computing intensity required to generate realistic
representations of the processing functions at a scale that is useful for
understanding changes in processing and rock interaction over longer time scales.

Output from constrained simulations that have been constructed and executed to
determine the ‘single best design’ cannot be used for risk mitigation evaluation ,
over and above the limited sensitivity analyses that have in some cases be used to
assess the impact of uncertainty (Nicholas, 2007).

3.3 The problem statement and research question

Current known techniques for calculating the expected metallurgical recovery of

diamonds for kimberlite diamond deposits are based on global assumptions and
average values and do not explicitly incorporate the spatial variation of rock
properties.

The problem stems from the following inadequacies:

74 Stephen Coward - April 2020

 Chapter 3: Problem statement

• Limitations of existing methods used to measure and estimate the physical

spatial characteristics of kimberlite;
• Mine production and metallurgical process modelling is not suited to
incorporating multiple multivariate spatial estimates of the physical
characteristics of kimberlites; and
• Project evaluation frameworks do not readily incorporate the change in
project value that can be attributed to mitigating recovery factor variability
and or uncertainty.

Through a combination of sampling, measurement, modelling, and simulation

experiments and two case studies, this research develops and demonstrates a
methodology to quantify the range of the metallurgical recovery factor for diamonds
recovered from kimberlitic deposits. This methodology improves the estimate of the
expected metallurgical recovery factor and provides an approach to derive
confidence limits.

A sampling experiment has been conducted on the K2 kimberlite at Venetia mine in

South Africa to generate a data set of measurements of rock characteristics.
Measurements were taken on cores, and the holes from which they were retrieved
to demonstrate methods to improve spatial data collection and show how these can
be used to generate spatial estimates of rock characteristics.

In parallel with this work, a modelling and simulation framework has been
developed to demonstrate how multivariate estimates and spatial simulations of the
physical characteristics of an orebody can be used in a linked-up value chain model
to derive a metallurgical recovery factor and create confidence limits for the range
and variability in the metallurgical recovery factor over time.

A quantified view of the variability and uncertainty of metallurgical performance

facilitates a quantitative evaluation of a range of risk mitigation strategies. These
strategies include blending and stockpiling, selection of alternative grinding and
crushing technologies, or adapting the processing configuration. The benefits are
quantified in monetary terms using simplified discounted cashflow metrics.

Stephen Coward - April 2020 75

 Evaluation of metallurgical recovery factors for diamonds recovered from kimberlites

4 OREBODY SAMPLING AND

COLLECTION OF ROCK
CHARACTERISTIC DATA

4.1 Introduction

The methodology used in this thesis to derive the metallurgical recovery requires a
spatial estimate of country rock and kimberlite characteristics that impact on the
diamond recovery process. The estimation of these characteristics is not trivial and
requires the collection and testing of samples to derive appropriate spatial data of
these characteristics. This chapter describes the design and execution of an
experiment to gather data that was used to evaluate options for generating spatial
estimates and simulations of the required kimberlite characteristics.

The nature of so-called ‘non-grade’ variables requires special consideration during

sampling and subsequent spatial modelling. Failing to objectively consider the
unique properties of these variables may seriously compromise the validity of the
sample selection, sample treatment, data models and the subsequent decisions that
are based on these models. In addition to addressing the vexed question of collecting
an ‘unbiased and representative’ sample, the sampling and modelling approach
must consider issues of scale, both geostatistical support and the relationship
between bench-scale data and operational-scale mineral processing performance
(Coward et al., 2009).

As identified in the literature review, the traditional approach to optimise process

design and treatment strategies is often based on the assumption that the waste, ore
and diamond characteristics are ‘consistent’ within a given kimberlite ‘facies.’ This
has meant that the sampling designs are often aimed at generating samples that will
be representative of the average characteristics of interest. These averages are then
used to derive recovery factors for each of the facies within the deposit. (Ashley and
Callow 2000). To create a sample that has these ‘average characteristics’ several
large representative samples are taken from the orebody and then mixed to form
what is called a ‘composite sample.’

76 Stephen Coward - April 2020

 Chapter 4: Orebody sampling and collection of rock characteristic data

There are several short comings of using a ‘compositing’ approach to sampling,

including:

• It is incorrectly assumed that an average of the results of measurements on

a number of composite samples will give the best estimate of the rock
characteristics that will be observed when the deposit is mined. As most rock
characteristics are not additive this approach could lead to incorrect results.

• An incorrect assumption that it is possible to compile a single sample which

has a suite of characteristics that are equivalent to those of an entire spatial
domain / facie /geological type. This is not possible as it would require prior
knowledge of the entire composition of the attributes throughout the
domain/facie or geology.

• A further incorrect assumption postulates that mixing and treating number

of different samples in equal mass proportions will result in the average
response being observed. If the mixing is not effective, and characteristic
being measured does not respond in a linear way to changes of rock
composition, a proportional blend will not give the equivalent proportional
change in the measured response. There are several reasons for this
including the difference in serialisation of the material; i.e. the mixed sample
will not have the same properties as the range of materials that are treated
during operations, as the properties will change in relation to the sequence
in which the ore blocks are mined.

• It is further assumed that it is possible to compile a range of composite

mixtures that will cover the range of expected mixtures that will be treated
over the life of the operation. It is also assumed that a physical blending of
the samples is possible and the response to the blend can be modelled in
some representative way. As an illustrative example, sampling is usually
limited in number by the required minimum sample mass to run steady state
tests (e.g., 100kg for a full drop weight characterisation test). To compile a
set of composites that covers mixtures of two rock types in increments of
10% one would need over a tonne of material, three rock types would
require several tonnes.

• The sample is not the estimate - A further shortcoming of the use of ‘average’
ore characteristics to determine recovery factors for a geological domain is
that the average of the results of treating the 'average composite' samples,
which are far smaller than the domain for which the predictions are being
made, do not constitute an estimate of the 'average recovery' that can be
expected for that domain, i.e. 𝐹(𝐸(𝑥)) ≠ 𝐸(𝐹(𝑥)).Without having sufficient

Stephen Coward - April 2020 77

 Evaluation of metallurgical recovery factors for diamonds recovered from kimberlites

sampling (where 'sufficient' suggests a large enough number of

representative samples of large enough support to enable the
characterisation of both the histogram and semi-variogram model in a way
that supports estimation) of the variables of interest it will not be possible to
estimate the characteristics at the right scale to reasonably predict plant
responses (efficiency throughput, balance etc.) at an operational scale. Using
averaged results, or responses from treatment of small samples as an
estimate or prediction of full-scale response is a flawed approach.

In the absence of project constraints (e.g., time, access, cost) sufficient samples of
appropriate size, i.e. geostatistical support, would be acquired from all domains in
the orebody so that both the average and variability of the characteristic variables
can be spatially modelled with a known degree of uncertainty. In addition to
collecting enough samples, the sampling programme will also consider the
relationship between bench-scale testing and operational-scale performance, and
the implications for the minimum mass of samples required for metallurgical tests.
(Coward et al., 2009)

Mineral projects are constrained by both time and cost, which in turn limits the
scope of sampling and testing programmes (e.g., limited access) that are designed to
characterise the orebody. This often leads to a trade-off between the number and
size of samples that are collected, measured and/or processed. Some tests and
measures require less sample mass, and for these it is possible to acquire large
numbers of samples e.g. point load tests (masses of several grams). The very small
support used to acquire data from the sample can compromise the measurement's
validity (i.e. the measurement does not correctly reflect the true value of the total
sample) and impact on overall sample representivity. It is important to be able to
quantify the impact of different numbers, sizes and types of samples on the scale-up
from sample support to production support. In cases where a large sample is
collected e.g., 100 tonne pit samples, the number of samples are limited, and so
spatial coverage is usually insufficient for estimation. Thus, it is often the case that
the support, spatial location and number of sample points are insufficient to
characterise the spatial character of the rock property variables and hence
rendering reliable spatial estimation difficult if not impossible.

This chapter describes the design and execution of a sampling campaign that aimed
to address many of the limitations of traditional sampling approaches. Several
different properties of kimberlite were measured using a range of methods and
tools on cores derived from holes that were relatively closely spaced (~5m).

It should be noted that this work was carried out in collaboration with Dr. Matthew
Field, who assisted with the design and management of the sampling campaign. The
work was generously funded by De Beers, who also allowed on-site staff at Venetia

78 Stephen Coward - April 2020

 Chapter 4: Orebody sampling and collection of rock characteristic data

Mine to supervise core drilling and core collection and logistics. The author co-
ordinated destructive tests on core samples at several laboratories. The author
undertook several site and laboratory visits during the process of the sampling
campaign. Rob Pierce assisted with project management and compilation the
geological data base.

The experiment objectives included:

• Assessing the relationships between the primary kimberlite properties

(geological description, contents) and the measured response variables
(UCS, BTS, acoustic velocity);

• Describing and evaluating the nature and strength of associations between

the different measurements on the core and down the hole;

• Identifying and using relationships between descriptive and measured

primary properties to assess the quality of prediction of characteristics of the
rock at sampled and un-sampled locations; and

• Defining the minimum range of descriptors and measurements that is

required to be able to characterise the mean, variance, variogram and
histogram of the variables of interest in the block of rock that has been
sampled.

The sampling design used in this experiment aimed to test the assumption that,
given sufficient geological controls (i.e. within a geologically homogenous, grade
‘domain’), it would be possible to sample kimberlite cores, and augment data
acquired from destructive rock testing with a range of geophysical measurements
to facilitate geostatistical characterisation and hence estimation of rock
characteristics. This required a data set to be generated from samples taken from a
real deposit that was geologically well defined, with closely spaced (i.e. less than 5m
apart) samples. Spatial intensity, or spatial coverage, can be defined as some
function of the average distance between the samples. Sufficiency criteria can range
from qualitative rules of thumb (Parker, 1978) for the mineralisation type to a
quantitatively derived measure such as having observations that are within half the
range of the semi-variogram. As no such data set was readily available for the
variables that were to be sampled in this kimberlite, this sampling experiment was
designed and executed based on the minimum feasible distance that could be
achieved within the physical constraints of the drilling equipment and the most
likely variogram range that would be expected for the variables considered.

The scope of the programme included the following activities:

Stephen Coward - April 2020 79

 Evaluation of metallurgical recovery factors for diamonds recovered from kimberlites

• Describe and classify the geological characteristics of the entire sampled

area;

• Design a layout of samples and core holes to acquire and test core from
several closely spaced drill holes of 50 m in length;

• Define how each portion of core is to be sampled and describe protocols for
core selection, preparation and treatment;

• Assess the operation of several downhole geophysical tools and potential to

use data generated from these tools to improve spatial estimates of rock
characteristics; and

• Investigate the correlations between different types of data acquired and

assess potential for easily measured variables (e.g. downhole geophysics) to
be used as reliable proxies for other test data that are more difficult or time
consuming to acquire (e.g., destructive rock strength testing).

This chapter describes the sampling layout, gives an overview of the geology of the
site selected, and describes the tests and tools used to generate the data. The
independent variables include all the primary properties of the rocks and the
dependant variables are the response variables. Process variables are those used to
control the properties of the processes used to carry out the tests.

A programme was carried out on a block of in situ ore (50m x 50m x 50m) located
at Venetia Mine in South Africa. Nineteen core holes were planned to be drilled and
each core was subject to several phases of description and testing according to a
pre-determined procedure.

The data set described in this chapter has been used in subsequent chapters to
demonstrate the proposed methodology to determine the impact that the variable
rock properties can have on diamond recovery.

This exploratory data analysis begins with an evaluation of the correlations between
measures:

• Between different types of destructive data;

• Between destructive and non-destructive testing;
• Downhole measures and destructive tests; and
• Quantitative Geological descriptors and physical rock responses.

80 Stephen Coward - April 2020

 Chapter 4: Orebody sampling and collection of rock characteristic data

In the next chapter, the spatial behaviour and correlations established between
these measurements are reviewed and their usefulness in generating spatial
estimates of the in situ the rock characteristics is assessed.

4.2 Experimental design

The main aspects of the specific sampling campaign that was carried out included:

• Site selection, hole layout and sub sample design;

• Non-destructive downhole and core measurement tools used;
• Destructive test selected;
• Data collection quality assurance and quality control measures applied; and
• Assumptions made in the in the absence of controls.

Location, hole spacing and sample layout

Experience from grade estimation exercises in kimberlites shows that diamond

grades commonly produce variograms with ranges of around 50m (Bush, 2010). It
was possible that the rock property variables that this programme investigated
would have a shorter range. Even if the range was larger than 50 m, the nugget effect
of these variables has not been documented in the known literature. Relatively short
5m centred drill hole spacing was thus selected for the experiment.

Orientated core holes were drilled at 5m centres from the centre hole in a cross
pattern to a depth of approximately 50m from the hole collar. This cross pattern has
been established as a reasonably robust methodology for establishing short-
distance component of variograms (Chiles et al., 1999). The reason for the close-
spacing of these holes is that one of the most important aspects of the spatial
relationships are the nugget effect (i.e. random component) and the range over
which individual measurements are correlated. If the expected range is 50m, then it
is essential that some data points are located between 0 and 50m.

The location of the cluster of holes was determined largely by the availability of
suitable areas in the open pit that would allow for access and extraction of the cores.
Mapping of the K2 open pit by Brown (2008) provided good geological control, and
the pit was made available for experimental drilling from November to December
2004. The cluster of holes straddled the faulted contact between “fine-grained,
homogenous volcaniclastic kimberlite” occupying the eastern part of the pipe, and
the more varied lithologies lying to the west of the faulted boundary. This approach
ensured that the data set would cover two distinct domains (Figure 13).

Stephen Coward - April 2020 81

 Evaluation of metallurgical recovery factors for diamonds recovered from kimberlites

Figure 13: Geological map of Venetia K2 (after Brown 2008) showing the location of the
Geomet drill holes (orange cross).

Two categories of drill holes were defined; The first are holes where all the material
was sampled, and the second where only portions were sampled.

Figure 14: A view of the layout of the core holes depicting the location of the subsamples.

The sampling layout of the cores extracted is depicted in Figure 14. White
intersections were not sampled but retained; all other coloured intersections were
sampled. Details are of the subsamples are shown in Figure 15. It was important to
assign a sample type to each of the lengths of core prior to recovery of the core to

82 Stephen Coward - April 2020

 Chapter 4: Orebody sampling and collection of rock characteristic data

minimise selection bias. This is of importance when selecting cores for destructive
testing. Post drilling selection methods can result in only competent cores being
selected for test work which introduces large biases in the resulting data.

Figure 15:Listing of subsamples taken from cores that were fully sampled.

The layout ensured that the fully sampled holes would provide a duplicate set of
samples for each mining bench, whilst the sub-sampled holes would provide a one
sample set per bench.

Ideally, all the data would be collocated at the same point and all the data would be
available at all points, and this data could then be termed ‘isotopic.’ As several of the
sampling tests destroy the core it is not possible to generate truly isotopic data. The
destructive tests were thus ‘bracketed’ with petrographic samples to test for short
scale changes in geological characteristics. The petrographic data were used during
analysis of the data to identify the bench sets where the assumption of geological
continuity within the bench would be tenuous.

A ‘sample set’ consisted of:

• three small petrographic samples (Pet) spread through the interval to enable
the geology of the section to be described;
• two microdiamond samples (Mida) to evaluate the diamond grade of the
sampled interval;
• a geometallurgical sample (Geomet) that would be used to determine the
physical rock characteristics; and
• a geotechnical sample (Geotech) that was used for destructive geotechnical
testing.

Stephen Coward - April 2020 83

 Evaluation of metallurgical recovery factors for diamonds recovered from kimberlites

Non-destructive core analysis

The core that was extracted was logged and subjected to a range of non-destructive
description, analysis and destructive tests. This included:

• Descriptive logging by an onsite geologist, to identify lithological type,

xenolith types, abundance and distribution;
• Images of the core were captured using a DMT (Deutsche Montan
Technologie GmbH) core scanner. This unit generated high-quality visual
spectrum light images of the core, creating a permanent record of the fresh
core. The images were also used to obtain modal analyses of the particles that
make up the rock, and to perform a structural analysis of the rock.
• Selected portions of the core were scanned using a Geotek core scanner. This
recorded core-diameter, P-wave, gamma-ray attenuation, magnetic
susceptibility, core-imaging, natural gamma and electrical resistivity.
• Selected portions of the core were also scanned with a near infrared spectral
scanner to generate images in the near infrared wavelengths. These images
can be analysed to infer the type and abundance of clay mineral minerals.
Hydrous minerals such as phlogopite, serpentine and smectite clays make up
most of the fine-grained matrix of kimberlites. These minerals play a role in
determining a kimberlite’s strength and slime content. As these minerals are
often very fine-grained, they are not easy to detect, and conventional analysis
by X-ray diffraction techniques is difficult, time-consuming and expensive. In
recent years infra-red spectrometry has been used as an exploration tool in
kimberlite exploration because it can detect the Mg-OH bonds present in
these minerals. Development of this system into a core-scanning technique
makes the acquisition relatively intensive spatial data of these mineral
species possible.

Downhole data collected

The downhole geophysical data were collected to evaluate the potential for building
relationships between several continuous measures and the discontinuous
destructive tests that were carried out on the core. These relationships can be used
to enhance the estimates of destructive test responses at unsampled locations. The
nature of this data also lends itself to assessment of the representativity of the sub-
samples that were taken from the core. As an example, the density distribution of
the sub-samples can be readily compared to that of the density measures derived
from the geophysical tools.

As these geophysical data are continuous, they can to some extent be used to assess
the impact of support on the derived measurements. The true support of the tools

84 Stephen Coward - April 2020

 Chapter 4: Orebody sampling and collection of rock characteristic data

not only differs from tool to tool but will also vary between rock types that respond
differently to the signals being used.

The downhole data were gathered using several tools, each of these tools and their
rock-tool responses are briefly described below:

Downhole Calliper

This tool measures the size of the hole using three arms that trail behind the tool
and rub up against the walls of the drilled-out cavity. The deflection of the arms is
measured and the circumference of the hole at that location is calculated assuming
that the hole is circular. Variations in the size of the hole are related to different
properties of the lithology that has been drilled through and to some extent the
drilling parameters. Using the calliper data, it is possible to calculate the volume of
the core extracted. Comparing the calipered volume against the measured
recovered core volume makes it possible to calculate in core recovery. Core recovery
is correlated with both rock quality and drill rig operating parameters.

Resistivity

The direct contact resistivity tools work on having several point contacts with the
wall rock that are used to pass a current through the wall of the hole. The rock
resistivity is determined from the relationship between the voltage that is applied
across the contacts and the current flow that is realised. Some tools induce a current
in the wall using coils that are energised with high frequency alternating currents
and induce a current in the wall rock. This induced current is inversely proportional
to the resistivity of the rock being tested. Resistivity is deduced from the amplitude
and relative phase of the secondary magnetic fields that are sensed by measuring
coils. Direct contact tools provide high resolution logs but do not provide the
penetration of the wall rock that can be achieved with inductive tools.

Natural gamma radiation

The natural gamma tool uses a sodium iodide scintillation detector to measure the
natural gamma ray radiation emitted by the formation that it is passing through. The
signal is processed into five distinct peaks that correspond to the radiation energy
from three most common elements that contribute to naturally occurring radiation;
potassium, thorium and uranium. The count rates in each of the five segments of the
energy spectrum are used to determine the relative proportion of these components
in the rock. In strata where relatively, low counts are obtained logging speed can be
reduced to ensure that the number of emissions measured remains statistically

Stephen Coward - April 2020 85

 Evaluation of metallurgical recovery factors for diamonds recovered from kimberlites

valid. The data can be expressed as total gamma ray, a uranium free gamma ray, and
the concentration of uranium, potassium and thorium.

Long and short spaced density

There are several tools that measure density down the hole. In this case the
instrument (or sonde) consisted of a gamma ray source that is contained in the
downhole probe. The gamma rays are scattered by the wall rocks. The intensity of
the returning gamma rays is a function of the electron density of the rock that in
turn is related to the bulk density of the rock. The back scatter is detected at two
distances from the source. The further the detector from the source the greater the
support of the density measurement. Short spaced readings may in some cases may
biased toward that of the drilling fluid if the hole surface is coated with drilling mud.

Formation hardness testing tool

This tool has been developed to measure the hardness of the wall rocks of boreholes.
The unit consists of a calliper that measures the original hole diameter and has a
cutter wheel that runs along the outside of the hole. The deflection of the cutter into
the wall is recorded at small intervals and reported as a penetration reading in mm.

Acoustic velocity

The velocity and attenuation of acoustic signals in rocks is related to the nature of
the material that is carrying the signal including its density, porosity, the fracture
frequency and orientation as well as the clast size distribution. The strength and
frequency of the input signal can be adapted, to improve the quality of the data
obtained.

There are several ways in which acoustic velocities in rock can be measured. Most
rely on transmitting an acoustic signal through a specimen of rock and measuring
the speed and attenuation of the signal of both compression waves and shear waves.

An alternative approach for extracted core is to create a right cylinder of the rock
and then input a signal of varying amplitude and frequency until the resonant
frequency of the specimen is obtained.

If it is assumed that that rocks being measured are homogenous, linear elastic and
propagate acoustically induced waves at the same velocity in all directions (
isotropic) it is possible to create a mathematical relationship between measured
shear wave and compression wave velocities and several of the physical moduli.
This approach is particularly useful for characterising the physical characteristics of

86 Stephen Coward - April 2020

 Chapter 4: Orebody sampling and collection of rock characteristic data

the rocks considered and so the relationships between acoustic velocities and
material characteristics are described here in more detail.

There are three broad categories of moduli:

1. Young’s modulus;
2. The Bulk modulus; and
3. The Shear modulus.

Young’s modulus is a measure of the elastic stiffness of a material. Named after the
18th Century British scientist Thomas Young. It describes how a solid deforms under
load within its elastic limit. It is defined as the ratio of imposed pressure to relative
deformation (strain) measured in a direction that is parallel to the direction of the
imposed pressure. Strain is a measure of deformation normalised to the sample
length or area (Equation 6). It can be derived from measuring the slope of the stress
strain curve near the origin, or in the interval from the onset of the deforming
pressure to a point before plastic deformation occurs. It is in this range that the
deformation on the solid is directly proportional to the imposed force. This
behaviour is described by Hooke’s law, so named after the 17th century British
Physicist, Robert Hooke.

∆𝑳
𝒆=
𝑳 Equation 6

Where

𝑒 is the calculated strain reported in dimensionless units;

∆𝐿 is the measured change in length of the sample; and

𝐿 is the original length of the sample.

Young’s modulus is described in Equation 7

𝑻𝒆𝒏𝒔𝒊𝒍𝒆 𝑺𝒕𝒓𝒆𝒔𝒔 𝑭⁄𝑨

𝑬 = 𝑻𝒆𝒏𝒔𝒊𝒍𝒆 𝑺𝒕𝒓𝒂𝒊𝒏 = ∆𝑳/𝑳 Equation 7

where

𝐺 is the derived shear modulus in Pascals (Pa);

Stephen Coward - April 2020 87

 Evaluation of metallurgical recovery factors for diamonds recovered from kimberlites

𝐹 is the applied force;

𝐴 is the area of the sample being tested;

∆𝑥 is the measured tangential displacement of the sample; and

𝐿 is the original length of the sample

The bulk modulus is the response of a solid to uniform multi-axial compression

forces, it is described in Equation 8.

∆𝑷
𝑲 = −𝑽 ∆𝑽 Equation 8

where

𝐾 is the derived Bulk modulus measured in Pa;

𝑉 is the volume of the sample;

∆𝑃 is the change in pressure applied to the sample;

∆𝑉 is the change in the volume observed for the given change in applied pressure.

The inverse of the bulk modulus gives the compressibility of the sample

The shear modulus, sometimes referred to as the modulus of rigidity, is the ability
of a sample of material to resist deformation in a direction that is tangential to the
main axis of the sample (Equation 9). It is reported in Pa and is measured by
applying a bending force to a sample and measuring the deflection as the bending
force is increased.

𝑭⁄𝑨
𝑮 = ∆𝒙/𝑳
Equation 9

where

𝐺 is the derived shear modulus in Pa;

𝐹 is the applied force;

88 Stephen Coward - April 2020

 Chapter 4: Orebody sampling and collection of rock characteristic data

𝐴 is the area of the sample being tested;

∆𝑥 is the measured tangential displacement of the sample; and

𝐿 is the original length of the sample.

It is possible to estimate the above moduli using measures of the acoustic velocity
in a solid. Alternative equations that can be used to estimate these quantities are
given in Table 6. These equations have specific measurement requirements, and are
based on assumption that the rocks being tested are linear elastic, homogenous and
isotropic (Momayez et al.,1995).

Young’s Modulus Shear Modulus

E
(1 −  )(1 − 2 ) G=
E =V 2
 2(1 +  )
1−
p

E = V p2 2 (1 +  ) G = Vs2 

9Gk
E=
(G + 3K )
E = 2G (1 +  )
Table 6: Formulae for estimating elastic moduli of solids using measured acoustic
velocities in rock specimens.

Where:

V P = compression wave velocity

Vs
= Shear wave velocity

 = density

E = Young’s modulus of elasticity

G = shear modulus of elasticity

k = Bulk modulus of elasticity

Stephen Coward - April 2020 89

 Evaluation of metallurgical recovery factors for diamonds recovered from kimberlites

 = Poisons ratio

Figure 16: Schematic representation of preparation and analysis of petrological samples.

(GSC=Geoscience Centre laboratory, UCT=University of Cape town, UOB=University of
Bristol, SGS=Testing laboratory in Johannesburg, Edinburgh =Edinburgh University).

Prior work on kimberlites has included an analysis of Venetia kimberlite material.

Measurement of the broadband acoustic properties of kimberlite used a wide range
of input frequencies, and compared the response of core and quarried specimens
(Sothcott et al., 2005). Six samples were tested at ultrasonic frequencies, and one
sample at range of sonic frequencies 3 kHz to 60 kHz.

The work showed that the ratio of compression wave velocity to shear wave velocity
was lower for samples taken from the pit than for core samples. Further analysis
indicates that this is a function of both weathering and potentially the introduction
of micro-cracks introduced during the mining process. The author did, however,
conclude that acoustic signal measurement could be used to discriminate between
waste rock species and different lithotypes in the Venetia deposit.

Destructive core testing

Once all the non-destructive testing and core description was completed, the
process of sampling the core to obtain further information commenced.

The following types of samples were taken from the cores:

90 Stephen Coward - April 2020

 Chapter 4: Orebody sampling and collection of rock characteristic data

Petrological: These samples were removed according the procedure shown above
and were submitted to the De Beers Geoscience Centre Johannesburg for sample
preparation according the scheme depicted in Figure 16.

Micro-diamond Samples: These samples were removed and submitted to the

Kimberley Acid Laboratory as per De Beers’ standard procedures. Diamonds were
recovered down to 150 microns.

Geometallurgical samples:

These samples of 1 metre of core were sawn to length, dried and packed in
cellophane and dispatched from site to SGS laboratories in Johannesburg.

~40 Discs Schematic of

Geometallurgical Sample
~ 25mm thick preparation and treatment
1 m of
63mmd 160 Triangles
ia core
quarter of core size

Drop weight
tester-
Screen t10 index

Ta Analysis TA index

Micro diamonds
Figure 17: Preparation and analysis of Geometallurgical samples.

The method used for rock breakage testing follows the procedure described by
Morell (Morell, 2003) and termed the SMC test. This test has been developed by the
JKMRC in conjunction with several industry partners. The method has recently been
refined by Dr S. Morrell to utilise a smaller support of core (Morell, 2003).
Essentially the method aims to impart a fixed energy into several fragments sawn

Stephen Coward - April 2020 91

 Evaluation of metallurgical recovery factors for diamonds recovered from kimberlites

from the core and the resultant degree of fragmentation is measured by sieving the
progeny of the breakage test. It is then possible to relate the energy input to
proportion of material that has been crushed to less than one tenth of the size of the
original fragments. This is the so-called t10 and gives a measure of the samples’
resistance to crushing.

At the laboratory in Johannesburg the cores were cut into quarters and then each
quarter section was cut into 40mm lengths to form triangular sectors of quarter core
for testing. Each of the triangles was then subjected to drop weight testing. The
product of the controlled energy fracture was collected and sized. These data were
then used to determine the energy breakage relationship for each sample (Figure
17).

Figure 18: Preparation and testing of Geotechnical samples.

Geotechnical Samples: A half a metre of core was sawn using kerosene as the cutting
fluid. The section of core was then cut into lengths of approximately 120mm for
uniaxial compressive strength testing. The remainder of the core was cut into discs
of 40 mm of core length for Brazilian tensile strength as described in the literature
review section. This process is depicted in Figure 18.

92 Stephen Coward - April 2020

 Chapter 4: Orebody sampling and collection of rock characteristic data

4.3 Data collection process

The samples were extracted from Venetia, a kimberlite mine located in northern
South Africa. 17 cores were extracted from the K2 pipe in the form of a cross with
the samples located 5m apart. The samples were collected from the core as per the
recipe described above.

The plan view layout of the core location is shown in Figure 19.

Figure 19: XY projection of the layout of the core samples.

The samples were collected in the middle of 2005 and subjected to destructive
testing shortly afterwards. Rocklabs, a testing laboratory located in Pretoria carried
out the UCS and BTS testing and SGS laboratories in Johannesburg carried out the
drop weight tests.

Cores were drilled out, collected and scanned using a visual spectrum core scanner.
The holes were subject to a suite of geophysical measures. The scanned core was
then marked according to the original sub-sampling design, regardless of the
‘suitability’ of the selected interval, this was done to minimise selection bias. The
sub-sampled lengths of core were also photographed to ensure that. where
anomalies were detected later in the analysis, the photographic record would
provide a means of further investigation.

The core was then cut into subsamples that were dispatched to several laboratories.
As the data were returned, they were compiled into an access database. Substantial
time was spent ensuring correct location of the data using several methods. This is
important because to test associations between measures (e.g. Downhole
geophysics and UCS) there must be a high resolution on the data location.

Stephen Coward - April 2020 93

 Evaluation of metallurgical recovery factors for diamonds recovered from kimberlites

4.4 Data acquired

The data collected are described using descriptive statistics that include some
measures of spatial dispersion and expected relationships in the data set. This
section provides data for this thesis and thus needs to be detailed in a way that
supports further analysis and identifies shortcomings in the data collection process
that might in future be improved by adapting the data acquisition protocol.

The data comprise several types of tests, collected from several different
laboratories, and have been stored in an access database titled Venetia K2.mdb. The
numbers of samples collected during this programme are given in Table 7.

Sample Type Number of Samples

Geomet 130
Geotech 130
MIDA 120
Petrological 627
Table 7: Summary of the number of each type of sample collected.

The aims of the descriptive statistics include:

• Describe the data that has been collected, identify any limitations in the data
collection process;
• Data preparation – removing outliers e.g., impact of casing on geophysical
readings;
• Evaluation of the relationship between geological characteristics and test
responses;
• A comparison and contrasting of the results from the different destructive
testing methods; and
• Identification and recommendation of changes improvements that can be
made to improve the reliability of the results of this type of testing.

The types of data collected included:

• Petrological logging
• Microdiamond sampling
• Downhole and core geophysical measurements
• Destructive test measures

94 Stephen Coward - April 2020

 Chapter 4: Orebody sampling and collection of rock characteristic data

Results of Petrological Sampling

A total of 43 petrological samples were acquired. The core sections were cut into
slabs and polished. The surfaces of the slabs were photographed at high resolution
and the images. The variables that were generated during this phase included the
grain size distribution and mineral abundance, the summary statistics of this data
are displayed in Table 8.

Total Lithics Olivines Matrix Lithics Olivines Matrix Lithic Olivine

Area Area Area Area Area% Area% Area % Grains Grains
mm 2 mm2 mm 2 mm 2 mm2 % % count count
Minimum 7938 2242.3 0.0 0.0 16.99 0.00 0.00 1 0
Median 15067 5690.6 1564.5 7757.2 36.99 10.43 50.51 1837 896
Mean 14846 6190.8 1519.6 7136.1 42.17 10.09 47.74 1887 1057
Max 16770 13314.9 2825.2 10870.1 100.00 17.80 68.10 4259 3132
Std dev 1699 2488.1 662.4 2407.5 17.82 4.00 14.72 1076 696
Coeff Variation 0.11 0.40 0.44 0.34 0.42 0.40 0.31 0.57 0.66
Count 43 43 43 43 43.00 43.00 43.00 43 43
Table 8: Summary statistics of analysis carried out on petrological slab samples.

The petrological data show the average proportion of area of each sample that is
made up of typical so-called ground mass minerals (matrix), olivine and lithics. The
olivine gives an indication of the proportion of deep mantle minerals (including
diamonds) that are in the sample, the matrix gives an indication of the state of the
kimberlite during emplacement, and the lithics give some indication of the
proportion of dilution in the kimberlite sample. For this data, although the matrix
makes up about half of the area sampled, in some specimens this value drops to zero
when the waste contamination rises to 100%. In this sample set, olivine constitutes
on average 10% of the area but it ranges from 0 to 17% in the samples.

Destructive test methods

Three different tests were used to determine rock fracture properties:

• Brazilian hardness test
• Uniaxial compressive stress
• Drop weight fracture testing
A brief description of each testing methodology follows:
Brazilian hardness testing:

The specification for this test is given by the American Society for Testing and
Materials (1988). It can be described as a compression to failure test where a core
specimen, cut into a disc of predefined dimension, is placed under load to failure.

Stephen Coward - April 2020 95

 Evaluation of metallurgical recovery factors for diamonds recovered from kimberlites

During the compression of the core the deformation of the sample translates the
compressive force into a tensile force that leads to tensile failure.

Uniaxial compressive strength

This is also defined as an ASTM test. In this case a section of core is cut to a
rectangular cylinder of predefined dimensions and loaded until failure. As the core
is loaded the deformation of the core can be measured by means of strain gauges
and several other derivatives of rock strength can be calculated.

UCS BTS
Specimen UCS Specimen BTS
Density Strength Density Strength
g/cm3 Mpa g/cm3 Mpa
Minimum 2.38 14.96 2.32 1.3
Median 2.52 31.11 2.52 5.1
Mean 2.56 44.66 2.56 6.2
Max 2.94 169.36 2.90 35.5
Std dev 0.13 39.49 0.14 4.0
Coeff. Variation 0.05 0.88 0.05 0.7
Count 137 137 135 136
Table 9: Summary statistics for the Uniaxial Compressive strength (UCS) and Brazilian
Tensile Strength (BTS) test work.

Drop weight tests

A total of 130 core samples that were 63mm in diameter were submitted for testing.
A total of 117 results were used for analysis. The method is described in Napier-
Munn et al. (1999) and requires several specimens of known size to be crushed with
a known energy input.

The ‘no data’ responses were a result of not being able to cut the core to the required
size. This suggests that these samples either were mishandled or had weathered to
an extent where cutting was made impractical. The infrared scans of these cores
were reviewed to identify high proportions clay in the following samples: 01, 02, 09,
20, 2, 26, 36, 44, 66, 68, 69, 96, 111, and 115.

The t10 data were gathered at three different energies and included the density and
size distribution of the product material after it was subject to drop weight testing.
A total of 117 results were obtained, and summarised statistics are given in Table
10.

96 Stephen Coward - April 2020

 Chapter 4: Orebody sampling and collection of rock characteristic data

Average t10 at 1 t10 at 0.5 t10 at 0.25

Average t10 at 1 t10 at 0.5 t10 at 0.25
Density Kwh/t Kwh/t Kwh/t
Density Kwh/t Kwh/t Kwh/t
33
g/cm
g/cm %
% passing
passing % % passing
passing % % passing
passing
Minimum 0.0 0.0 0 0
Median 2.5 32.8 20.89 10.75
Mean 2.3 29.9 18.86 9.83
Max 2.8 41.5 27.04 15.29
Std dev 0.7 10.7 6.92 3.97
Coeff Variation 0.32 0.4 0.37 0.40
Count 118 117 117 117
Table 10: Summary statistics of drop weight results obtained from cores.

In Table 10 it is noted that the coefficient of variation of the t10 increases with
decreasing energy. This large change in the range of the variability of the response
variable may be the result of a few outliers in the data set.

It appears that at an energy input of 0.25 Kwh/t, at which only 10% of the material
has been reduced to one-tenth of its original size, that the differences in the pre-
existing fractures of the rock specimen used for drop weight measurements has
more of an impact than the intrinsic rock fabric. Histograms of this data are
presented below.

Figure 20: Histogram of the density of samples subjected to drop weight testing.

The t10 data range from 41 to 4% passing one-tenth of its original size at three
different energy levels. There are some samples which were identified as waste rock
that gave results that lie in the tails of country rock; these have been labelled as
waste rock.

It is however known that there is a relationship between the lithology from which
the sample comes, the density of the sample and the resulting susceptibility to
fragmentation. The histograms in Figure 21 show the spread of this data. In this
data, the increasing spread of responses is also indicated (the top panel in this figure
is at a lower energy than the histogram in the lowest panel). This suggests that at
higher energy input levels the range of resulting fracture is higher, and the

Stephen Coward - April 2020 97

 Evaluation of metallurgical recovery factors for diamonds recovered from kimberlites

histogram is less skewed. This gives an important indication to the design of future
experiments, suggesting that appropriate energy levels are required to fully
characterise the rocks energy breakage relationship.

Figure 21: Comparison of histograms of drop weight results for all samples tested at three
different input energies (top figure is lowest energy; bottom panel is highest energy).

98 Stephen Coward - April 2020

 Chapter 4: Orebody sampling and collection of rock characteristic data

The samples were taken from two different rock types, a breccia domain which
includes many country rock fragments, and a more coherent volcaniclastic domain.
The data can thus be grouped using the rock type criteria. Figure 22 and Figure 23
present a comparison of the location and histograms of density for the VKBR and VK
domains respectively

Figure 22: Histogram and base map of sample density for the VKBR domain.

From these figures it is apparent that the VK domain has a smaller variation of
density and so we would expect to see a less variable measure of rock fracture for a
given energy input as depicted in Figure 24.

Figure 23: A histogram and cross section of the densities of the samples taken from the VK
facies domain

As confirmed by the lower dispersion of the histograms depicted in Figure 24.

Stephen Coward - April 2020 99

 Evaluation of metallurgical recovery factors for diamonds recovered from kimberlites

Figure 24: Comparative histograms of drop weight responses at three different input energy levels
for the VKBR facies (LHS) and the VK facies (RHS).

This change in location and dispersion of the t10 variable with differing energy input
suggests that during analysis modelling of this variable in a spatial context will not
be trivial. Several approaches to dealing with the skewness and instability of the
histogram will always require testing if this variable is to be used.

100 Stephen Coward - April 2020

 Chapter 4: Orebody sampling and collection of rock characteristic data

Descriptive statistics of downhole wireline geophysical data

The downhole data were acquired by running several tools down and up the drilled
holes at a controlled rate. All data were then provided to the project and
summarised in a database.

Table 11 below contains summary statistics of the results for all holes that were
surveyed using the tools provided.

Summary Statistics- Data obtained from Wireline Geophysical measurments

Short Long Magnetic
Measured Clipered Spaced Spaced Suceptabilit Neutron Natural P Wave S Wave
Depth diameter Density Density y Count Gamma Resistivity Velocity Velocity
m mm g/cm 3 g/cm 3 mm2 # Api units Ohm/m2 m/s m/s
Minimum 0.3 0.0 0 0 -11 0 0.0 0.0 0.0 0.0
Median 24.6 96.3 2.58 2.57 0.3 714.1 90.2 9.9 263.9 488.4
Mean 29.1 95.1 2.53 2.50 17.4 726.1 89.4 34.0 246.6 464.1
Max 69.3 123.2 3.71 3.54 831.0 1698.2 210.7 5013.2 323.4 617.7
Std dev 19.7 13.0 0.58 0.53 110 163 27.1 278.6 58.3 106.1
Coeff Variation 0.68 0.1 0.23 0.21 6.34 0.22 0.3 8.2 0.24 0.23
Count 9726 9726 9726 9726 9726 9726 9726 9726 9726 9726

Table 11: Summary of downhole geophysical readings taken from each hole drilled.

A brief analysis of this data indicates that there is an impact of the steel casing that
was used around the collar. This can be seen in the histograms for density which
shows several observations in the 3.5 g/cm3 range (Figure 25). The main
implication is that the observations from the cased portion of holes can be used to
check the calibration of the density measures, and that it is important to discard this
information from any further rock property interpretation.

Figure 25:Histogram of “short spaced density” showing high readings.

Stephen Coward - April 2020 101

 Evaluation of metallurgical recovery factors for diamonds recovered from kimberlites

Statistics of downhole formation hardness testing

The analysis of the data produced in this way aims to determine the following
relationships:
• Penetration and lithological and petrological classification;
• Penetration and destructive tests; and
• Penetration and other geophysical measures
The tool was deployed down four holes DDH 357, 358, 359, 365 and 366. The basic
descriptive statistics of the downhole formation data is contained in Table 12.

Hole Hole Hole Hole Hole

DDH357 DDH358 DDH359 DDH365 DDH366
Average 1.0 0.5 0.6 1.7 0.6
Maximum 17.8 4.9 15.7 14.7 11.5
Minimum 0.0 -0.6 -1.9 -0.4 -0.8
Standard Deviation 0.7 0.5 0.7 2.1 0.5
Coeff Variation 68.2 113.3 134.6 125.0 82.7
25 th percentile 0.7 0.2 0.3 0.4 0.4
75th percentile 1.0 0.6 0.6 1.8 0.8
Interquartile range 0.4 0.4 0.3 1.4 0.4
Table 12: Descriptive statistics of the downhole formation tester results.

A plot of the three readings for hole DDH357 is shown in Figure 26, a histogram and
downhole variogram of the cutter penetration (the difference in location between
the calliper and cutting wheel) is provided in Figure 27. The ranges of the data
clearly demonstrate that the unit is indeed responding to the different lithology
types and higher variances in penetration are seen in breccias. It is also interesting
to note that the points of extreme deflection are associated with break out from the
drilled cavity. This information can be used to quality control the downhole
calipering of the holes.

Raw Formation Hardness Data- Hole 357

100 5
99 4.5
Difference in Deflection (mm)

98 4
97 3.5
MM diameter

96 3
95 2.5
94 2
93 1.5
92 1
91 0.5
90 0
10 20 30 40 50 60 70
Depth (m)

Caliper 1 Arm Hardness Caliper Dif ference

Figure 26: A plot showing the three readings generated by the formation hardness tool for
hole DDH357.

102 Stephen Coward - April 2020

 Chapter 4: Orebody sampling and collection of rock characteristic data

As can be seen from the downhole semi-variogram (Figure 27)the range of the raw
semi-variogram is of the order of 15 m, justifying the requirement for spacing of
holes by approximately 7m.

Figure 27: Histogram (LHS) and semi-variogram (RHS) of mm of deflection

('Difference')measured by the downhole formation testing device.

4.5 Observations arising from the data collection

The design and execution of this programme required substantial support from the
mine and its operational staff for which the researcher is very grateful. Ensuring that
the right equipment and access could be provided to the site required substantial
planning and coordination. It is essential that in the execution of this type of study
that frequent and on-going communication is facilitated.

Several of the test units, which although they were housed in an air-conditioned
container, had occasional problems with operating temperature and dust. As this
campaign aimed to get many different types of close spaced data from the samples
generated the programme took several months to complete. The study does
however demonstrate that it is possible to obtain collocated data of different types
for the evaluation of kimberlitic deposits.

The petrographic data that have been gathered provide a basis for segmenting the
geology of the test area based on the composition of the rock mass and the
abundance of different types of rock fragments. The analysis shows that one of the
limitations in the data collection was the misinterpretation of the location of the
contact between the breccia and coherent volcaniclastic kimberlite. This led to the
coherent volcaniclastic kimberlite being under-represented in the dataset.

The destructive data that have been gathered is co-located with multiple
geophysical measures and hence provides a basis for investigating the potential to

Stephen Coward - April 2020 103

 Evaluation of metallurgical recovery factors for diamonds recovered from kimberlites

use the geophysical data to augment the destructive test data and improve the
estimation of the physical characteristics of the kimberlite properties in the test
area.

The sampling programme required substantial supervision of the core extraction,

core handling and tracking through the protocol. Use of core and hole imaging
provided an excellent basis for quality control of the multiple processes that the core
was exposed to. As the two images can be compared side by side it is possible to
correctly allocate core loss and identify where cores have been incorrectly loaded
into core trays. Ideally this analysis should be carried out between core collection
and subsampling to prevent incorrect allocation of core to test destinations. The
primary use of the images in this experiment was to ensure that the geophysical data
that was collected was correctly matched with to the samples selected for
destructive testing.

4.6 Conclusions arising from data collection

A data set was generated for this research by samples taken from, and
measurements made on, two distinctly different rock types; a highly variable
kimberlite breccia and a more spatially continuous volcaniclastic lithology. The data
have been quality controlled and then captured and compiled into a substantial and
accessible database to facilitate further analysis.

This experiment aimed to demonstrate that it is possible to overcome some of the

challenges that are faced in the collection of samples and generation of data that can
be used to generate spatial characteristics of kimberlite rocks. Due to constrained
access to the orebody and limited sample mass that is typically acquired the sample
design must try, as far as possible, to optimise:

• the spatial location of the cores,

• the sub-sampling of the core and assignment of appropriate measures and
destructive tests on the core, and
• the use of the core hole that is drilled for additional data acquisition.

The next chapter expands on the statistics and reviews associations that exist
between the measures made and demonstrates how these can be used to generate
estimates of rock properties at un-sampled locations.

The work described here comprises a cost-effective method of generating unbiased

samples that can be used to measure a range rock properties. The colocation of
destructive and non-destructive measures facilitates the investigation of
relationships that can be used to predict the destructive measures at unsampled

104 Stephen Coward - April 2020

 Chapter 4: Orebody sampling and collection of rock characteristic data

locations. The data gathered here and their use in the value chain model to predict
the expected range of the metallurgical recovery factor is explored in the next
chapter

The quantitative understanding of the spatial characteristics of geological,

mineralogical, geochemical and geophysical variables and their impact on physical
characterisation can be used to improve the design of future sampling campaigns.

The approach that has been demonstrated here is applicable in the exploration
phase of new deposits and is also of significant use in assisting the sampling
programme design when extension of the resource and reserves of existing
operations is required.

Stephen Coward - April 2020 105

 Evaluation of metallurgical recovery factors for diamonds recovered from kimberlites

5 MODELLING PHYSICAL
CHARACTERISTICS OF
OREBODIES

5.1 Introduction

A model of the in situ ore and waste-rock properties that impact on the mining and
recovery of diamonds is required as an input into the value chain model that is used
to estimate the expected overall diamond recovery.

The properties that are important can be usefully classified as two types of
variables:

1 -Primary (in situ) variables: these are mostly characteristics that can be directly
measured, for example measurements of grade, density, grain size; and

2 -Response variables: these are measurements of the result of doing something to

the rock - for example recovery, throughput etc.

This framework is designed to assist with developing sampling approaches and

identifying the most appropriate spatial modelling approach. The proposed
framework can also help identify the risks associated with designing, sampling and
modelling of both types of geometallurgical variables (Coward et al., 2009). The
important aspect of the framework is that primary variables are, in general, far
easier to sample, and estimate, whereas response variables require far more
thorough consideration.

Current approaches reviewed to date that are used to model rock characteristics are
limited in that they aim to produce BLUE (Best Linear Unbiased Estimates) of the
characteristics. These properties of the estimate can be useful for several purposes
e.g., when the estimate is being used in a long-term design optimisation process but
may have some limitations when used in a value chain model as the estimates
produced are smoother than the real in situ variability. The methodology
demonstrated here suggests the use of estimates and simulations of the rock

106 Stephen Coward - April 2020

 Chapter 5: Modelling physical characteristics of orebodies

characteristics, which allows the variability and uncertainty of the characteristics to

be propagated through the value chain.

In this chapter the different methods of spatially estimating rock properties at an

appropriate scale for inclusion into a spatial block model are explored and their
impacts on the value chain model analysed. This leads to a recommended sequence
and approach to generate orebody models that have properties that will have the
required characteristics for use in the value chain model to derive the range of
values for the metallurgical recovery factor.

The process of building appropriate robust spatial models of rock properties

includes:

• Identifying and selecting the characteristics of the ore and waste rocks that
have a material impact on the process rate and efficiency;

• Design and execution of a sampling programme to acquire enough mass and

number of representative samples of material from each ‘domain;’

• Selection and preparation of samples for laboratory-scale treatment test work;

• Quality control of numerous laboratory tests;

• Relating primary physical rock characteristics to test responses;

• Support and scale corrections for laboratory-scale test results to process scale
response;

• Modelling of spatial behaviour of acquired data;

• Estimation and/or simulation of characteristics at block scale; and

• Evaluating the error, variability and uncertainty of the estimates made.

Figure 28 presents a schematic of an approach to developing models of required

rock characteristics.

The process of building spatial models is centred on data that can be acquired in
many ways, one useful distinction is that between ‘direct’ and ‘indirect’ data (Dowd
and Pardo-Igúzquiza, 2006). ‘Direct data’ are data acquired from a sample taken
from a known location. The sample data are acquired by subjecting the sample to a

Stephen Coward - April 2020 107

 Evaluation of metallurgical recovery factors for diamonds recovered from kimberlites

controlled measurement or testing process. This is referred to as direct in that the

result and the location, and to some extent the error, of the data can be directly
associated with the sample that has been treated.

Model
Prepared

Non Linear - Point Scale Post Populate

Complex Pathway Simulations Processing Spatial Model

Transform Not
Additive

Linear Transform Linear

Transforms Additive
Estimate
Means
Important
Not Additive

Assess Data Test for Select Important

Coverage Additivity Additive Data Characteristic

Low Coverage
Identify Variables Extremes
to be Modelled Important
Augment with Conditional
Meta data Simulation
Begin

Figure 28: Schematic showing pathways of generating spatial models of rock

characteristics.

The term ‘indirect data,’ on the other hand, refers to the acquisition of data where
the input is acquired from a tool that measures the rock response to a geophysical
input; the geophysical response is then used to inform the property of the rock that
would have exhibited had it been exposed to direct testing and measurement, but in
an indirect way though some form of mathematical model. The value of integrating
these two data types is that the relationship between the direct and indirect data
can be used to improve the estimate of the unknown quantity at un-sampled
locations.

Importantly, direct data are usually sparse and need to be expanded in coverage and
support. Indirect data can, however, be adapted to provide larger support, is far
more exhaustive and usually needs some form of compression or summarisation.

In this chapter, methods to improve the quality of the rock characteristic estimates
using relationships between the direct, indirect, non-destructive and destructive
data are explored. These improvements include the reduction of the error of the
estimate of rock characteristics derived from destructive tests at un-sampled

108 Stephen Coward - April 2020

 Chapter 5: Modelling physical characteristics of orebodies

locations and the identification of potential sampling bias that might arise from
sample selection.

The approach adopted can broadly be separated into two phases. The first is an
exploratory data analysis phase in which relationships between the variables of
interest are evaluated, followed by a second phase that evaluates different methods
for spatially estimating values into the block model needs to be considered.

This chapter begins with a brief review of the data that have been collected from the
experiment. These data are then used to a) improve the estimates of the rock
characteristics at point scale ‘down the hole’ using a few multivariate techniques
and b) to review the implications of different pathways that are available to
generate block scale estimates of response variables.

5.2 Estimation options considered and compared

Previous work on the integration of different sources of data has included several
methods that use spatial data to generate property estimates at block model scale
(Dowd 1997, Dowd and Pardo-Igúzquiza, 2006). These can be grouped into two
types:

1)Those designed to deal with large amounts of data in a reductionist way -

these use a variety of multivariate statistical methods, including linear
regression, principal components; and

2)Those used to understand and exploit the spatial nature of the data to
generate spatial estimates of variables at a mining block scale. These
methods also use spatial covariance models to integrate direct data and
sensed data .

Multivariate modelling techniques include linear regression modelling, principal

components analysis, and partial least squares models (Esbensen, 2002 and
Wackernagel, 1995).

Spatial modelling approaches considered here include polygonal estimation and

linear geostatistical methods. In Dowd and Pardo-Igúzquiza (2006) direct and
indirect sample data are integrated using several methods. The performance of the
various estimation methods was assessed by a range of statistics including mean
error and mean squared error.

The methods used included ordinary kriging, standard linear regression, and
estimation of a local mean, and simple kriging with estimation of a local mean,

Stephen Coward - April 2020 109

 Evaluation of metallurgical recovery factors for diamonds recovered from kimberlites

kriging with external drift, co-kriging, and Bayesian integration. The outcome of this
study indicated that if there is less than 10% direct data available, and the
correlation between direct and indirect data is less than 0.3 then geostatistical
methods are better as they give lower mean squared errors than linear regression.
Co-kriging is best when the correlation co-coefficient is above 0.2. Above 0.3
external drift and or local means perform best (Dowd and Pardo-Igúzquiza, 2006).

Spatial variables assume values that change as some function of the support and
locations at which they are measured. Most spatial estimation methods create some
form of weighted average the measured values of a variable at a point, over a specific
volume (e.g., the grade of a drill core is the average grade over a cylinder of rock).
The variance of a spatially correlated variable usually decreases as the support on
which the measurement is taken increases. There are several methods available for
using point data acquired from some location in space to estimate the value of a
spatial variable at an un-sampled location and these are briefly reviewed.

The simplest of these is to create a polygonal estimate where the values that have
been measured in the sample are used to inform a volume around the sample. In this
approach, no averaging is carried out, but neither is a specific change of support
calculation, so that the estimate will have a variance that is a function of the sample
layout, and hence volume estimated, and it is likely that the estimates generated in
this way will have a lower variance than the real values.

Ordinary kriging is a minimum variance, linear, unbiased estimator (Journel and

Huijbrechts, 1978). It has been applied successfully in many fields where the
variable of interest has a spatial component and is widely used for grade estimation
in mining. The estimate at any given location is a linear weighted combination of a
selection of the sampled values.

When using a kriging approach for estimation, the spatial variability of the data is
characterised in the form of a spatial covariance model or semi-variogram. Reliable
estimation of the parameters of the spatial covariance model requires sufficient data
that are representative of the domain within the orebody that is to be estimated.
Domains are variable-specific and selecting and defining domains of stationarity for
response variables requires consideration of both statistical and geological aspects
of the variable to be modelled.

Rock characteristic variables may require sampling on various sample volumes and
over a range of distances to provide an adequate quantification of spatial variability.
Linear geostatistical estimation methods (e.g. ordinary kriging) require variables to
be additive. Thus, if the variables are non-additive this process is not applicable but
may be used if a data transform can be used to render the data sufficiently additive,

110 Stephen Coward - April 2020

 Chapter 5: Modelling physical characteristics of orebodies

or if the range over which the variable is observed appears not to show too much
non-linear behaviour.

In some cases, however it is the extreme values that are of interest in the orebody –
e.g., for example extreme hardness that will damage crushing equipment. In this
case geostatistical simulation methods are required. These can be used to honour
the data at the sampled location and produce the variability at a scale suitable for
the value chain model. Individually the simulations present a ‘truer’ reflection of
variability in the ore characteristics, and when used as a set can give an indication
of the range of values for the characteristic of interest. Methods described in the
literature review will be demonstrated here.

5.3 Selection and preparation of physical characteristic data for

estimation

The variables selected for modelling are useful for predicting process performance
and diamond recovery. These broadly cover variables that will influence the
fracturing of the rock in a process, and the efficiency of the separation of diamonds
from their host rocks.

There are several destructive tests that can be used to measure the strength of rock
specimens. In this work the focus has been on obtaining data that quantify the
nature of the rock strength in tension and compression, and the fracture that results
from energy input. The tests considered include uniaxial compressive strength
(UCS), the Brazilian tensile strength (BTS) and t10.

Some existing test standards have well-defined support, even though this may be
inappropriate for all rock textures (ASTM, 1998). This is because the scale at which
they are carried out may not be sufficient to measure the responses that are valid
at the scale of the required estimates. This is arguably true of rock strength
measurement in coarse grained rocks or in breccia lithotypes. To overcome this, it
is suggested that several tests are conducted over a range of sizes larger and smaller
than those laid down in the standards in use.

Several of the rock characterising tests (Bond Work Index, Drop weight test, UCS,
BTS) have been developed to mimic at a small scale the processes that the rock will
undergo during mining and treatment. The underlying philosophy has been that if
one can obtain a representative sample it is possible to carry out small-scale tests
with sufficient rigour and control to enable their results to be scaled up using some
form of factorisation.

Stephen Coward - April 2020 111

 Evaluation of metallurgical recovery factors for diamonds recovered from kimberlites

5.4 Using multivariate techniques in estimation

During the acquisition of rock characteristic data, it is possible to augment the

primary point data with higher coverage, but potentially lower quality, geophysical
data. As the support of the destructive and non-destructive tests may be vastly
different, the raw data require some transformation to enable the two types of data
to be used together.

Scaling of geophysical data

Downhole geophysical readings create a valuable data set as the readings are
continuous and data acquired from different tools can be used in concert to improve
the estimate of destructive results at un-sampled locations. The downhole
geophysical wireline log readings are taken approximately every 10mm. At this
scale, the tools are potentially responding to several very small features, and the
readings may to some extent be smeared.

Bulk modulus vs Scale of measure

within and between sample variance

10
.

9
8
Average Variance in readings

7
6
5
4
3
Average variance within samples
2 Variance between samples
1 Total variance

0
0 1 2 3 4 5 6
Meters of reading averaged

Figure 29: A plot showing the impact of the size of the interval used on the total variance
of the bulk modulus measurement.

To improve the relationship between the wireline geophysical signals and the
destructive tests the geophysical signals need to be accumulated. The data were
accumulated over increasing intervals and the within and between interval variance
calculated. The interval over which the sum of these variances is minimised is
deemed to be the appropriate length for averaging the geophysical signal. In the case
of acoustic data this was found to be about 70cm of readings (Figure 29).

112 Stephen Coward - April 2020

 Chapter 5: Modelling physical characteristics of orebodies

This has the impact of reducing the signal variance and improving the correlations
between the wireline geophysical log and the results of the destructive tests. It is
strongly recommended that in the search for appropriate proxies this approach to
upscaling be used to evaluate associations between destructive and non-destructive
rock measurements.

Linear regression models

The correlations between all measured data were calculated. Some of the higher
correlations were observed between acoustic velocity and various measures of rock
hardness. Indeed, it is possible to translate acoustic velocities into various estimates
of destructive rock properties using the formulae discussed in Momayez et al.
(2004).

In this case it can be demonstrated that in the volcaniclastic facies there is a

reasonable correlation between the formation hardness tool estimate and the
measured UCS, as shown in Figure 30. In this plot, the correlation is relatively low,
even though several outliers have been removed from the data. This is partially
because of the breccia lithology being a very heterogeneous rock type.

Relationship UCS Vs Formation Hardness- hole 357

40
UCS_spec_strength
35 2012VK Linear model
2011VKBr
2014VKBr
2013VKBr
30
2010VKBr
2007VKBr 2008VKBr
.

25

2009VKBr 2006VKBr
UCS in Mpa

20

15

10

0
0.400 0.600 0.800 1.000 1.200 1.400 1.600
MM deflection of Cutter

Figure 30: A plot showing the linear model developed between the formation hardness tool
readings and the UCS values.

Even though the relationship is weak, it is possible to use these data in several ways,
such as co kriging to improve the estimate of the destructive variable (UCS) using
the downhole response (Cutter deflection).

Stephen Coward - April 2020 113

 Evaluation of metallurgical recovery factors for diamonds recovered from kimberlites

Multivariate techniques for data augmentation

There are several multivariate techniques available to integrate data of different

types. This creates an opportunity to augment sparse destructive data with more
abundant geophysical data to improve the estimate of the response properties.
(Dowd, 1997).

Figure 31: A box and whisker plot showing the sample value used to calibrate the PLS
models on the left of the plot, and the estimate of UCS down the hole on the right. The
heights of the bars indicate ‘goodness of fit’ of the partial least squares model derived from
multiple versions of the model using different combinations of sample and hold out data.

The sampling experiment gathered abundant geophysical data that are correlated,
albeit weakly with the spatially sparse destructive data. In this case a trial was made
using a PLS model to estimate the UCS of core along its length. To do this, a
calibration set of samples is required to define the relationship, which can then be
used to estimate the values at un-sampled locations. In Figure 31, generated using
Camo’s Unscrambler software, the sample values are plotted next to an estimate of
the values downhole 358. The very high bar on the left shows that there is one
sample (Venucs 2016) for which the relationship between the model and the actual
value is very weak and does not correspond with the model parameters.

There is also a way to incorporate directly the results of the acoustic velocity data
into the downhole estimates. These can be plotted together to provide a means of
visually reviewing the models as shown in Figure 32.

114 Stephen Coward - April 2020

 Chapter 5: Modelling physical characteristics of orebodies

Samples and Models of UCS - Hole 358

50 4

45
3.5

.
40
3
UCS values MPa

35 VENUCS2049 VENUCS2053
VENUCS2048 VENUCS2051
VENUCS2050 VENUCS2055 2.5
30

mm penetration
VENUCS2046 VENUCS2052
25 2
VENUCS2047
20
1.5
15
1
10
0.5
5

0 0
0 10 20 30 40 50 60 70
M Down hole

UCS Samples Modelled on FH Moddled on AT Acoustic signal Cutter Penetration

Figure 32: A plot of measured UCS values, and models for the samples based on the
downhole acoustic signal and cutter penetration depth.

In Figure 33 the UCS values down the hole for the samples recovered from the VK
facie have been calculated using a PCA model of P-wave acoustic velocity
measurements, S wave acoustic velocity measurements and measured densities.
The boxes show the range of values that would be calculated for redoing the PCA
multiple times and leaving out approximately 10% of the data in draw – this gives a
sense of the uncertainty or potential error that is in the PCA model.

Multivariate model validation and future work

Results of several methods have been presented, however rigorous testing of the
assumptions and reliability of each approach is still required. The ideal way to do
this is to generate simulations that can be ‘virtually’ sampled, and the samples used
in the different ways demonstrated here. This would provide a method to compare
and rank both the benefits and deficiencies of each method.

The second way in which these methods can be evaluated is to generate estimates
and then evaluate the predictions based on the actual properties that are
encountered when this part of the orebody is mined. The challenge in this approach
is that the estimated ore will not be the only source of rock to be treated. There will,
however, be an opportunity to carry out in-pit sampling as the mine deepens and

Stephen Coward - April 2020 115

 Evaluation of metallurgical recovery factors for diamonds recovered from kimberlites

Figure 33:PCA model of UCS downhole based on a Principal Component Analysis (PCA)
model using P- wave velocity, S wave velocity and long density, using one principal
component (upper panel) and two principal components (lower panel).

5.5 Estimation and simulation methods

The benefits and limitations of a few approaches to estimating rock characteristics

are explained and demonstrated.

Polygonal

For a highly heterogeneous domain in which the underlying spatial structure cannot
be discerned it is reasonable to assign each sample attribute value as the best
estimate of the attribute of a volume, of specified shape and size, centred on the
sample. In this case downhole models would be used to generate estimates for
samples of 70cm cores (as indicated by the prior work on upscaling). These can then
be used to estimate the values into polygons measuring 0.7m*0.7m *0.7m. In this
way, the sample statistics and the block statistics would be the same (which, of

116 Stephen Coward - April 2020

 Chapter 5: Modelling physical characteristics of orebodies

course, is impossible). The sample and estimate statistics using this method on the
experimental area are contained in Table 13.

Polygonal
Polygonal Block
estimate
Estimate values
Statistic Sample Values averaged into 5m
(0.7mx0.7mx0.7m
x 5m x 5m block
grid)
model

Count 116 376 094 1330

Minimum 16.02 16.02 21.2

Mean 33.30 33.15 33.42

Maximum 41.51 41.51 41.51

Variance 16.03 16.09 13.61

Table 13: Summary Statistics for drop weight test data showing sample and polygonal
statistics.

The underlying assumption with this approach is that the samples and modelled
values are representative of the areas between holes. When drilling is not closely
spaced this assumption may become very tenuous. The impact of sample layout on
the estimate generated using this approach is demonstrated in Figure 34. Note how
samples that are located far from other samples influence a far larger area than
samples close together.

Stephen Coward - April 2020 117

 Evaluation of metallurgical recovery factors for diamonds recovered from kimberlites

Polygonal Estimate of t10 values

Figure 34: A 3d representation of a polygonal estimate of drop weight values in the 0.7m
x0.7m x 0.7m grid.

Geostatistical estimation and simulation methods

The data were imported into Isatis, a geostatistical software package, and raw semi-
variograms were calculated, modelled and interpreted. Initial analysis has detected
ranges in the order of 6m to 15 m. This short range is thought to be due to the nature
of the orebody sampled and the small specimen size.

Increasing the support by ‘compositing down the hole’ has not decreased the overall
variance as the samples are not contiguous along the core. Further experimental
work would be required to determine the impact on both the distribution of
destructive rock property values and their ranges. One approach might be to acquire
larger diameter core and thereby increase the sample support. This would, however.
increase the cost and time taken for sampling.

The experimental and fitted semi-variograms for the UCS data are displayed in
Figure 35.

118 Stephen Coward - April 2020

 Chapter 5: Modelling physical characteristics of orebodies

Figure 35: An experimental and fitted model for the variogram for the UCS data.

The fitted semi-variogram model comprises a nugget effect and two spherical
structures with ranges of 25m and 27m. The kriging was carried out with an
isotropic search radius of 27m and used a minimum of 5 samples and a maximum
of 10. A cross-section of the area estimated is depicted in Figure 36, where hotter
colours represent higher values and cooler colours represent lower values.

Figure 36: N-S cross section of area estimated showing high UCS estimates in hotter colours
and lower values in cooler colours (LHS) and a histogram of the block UCS values estimated.

Stephen Coward - April 2020 119

 Evaluation of metallurgical recovery factors for diamonds recovered from kimberlites

The statistics of the estimates as shown on the histogram indicate that the mean of
the UCS is slightly lower than the mean of the samples but that the standard
deviation of the block UCS has dropped from 27 to 6.5. It is also evident that the
estimate in this specific case is influenced by a few very high values.

Fortunately, the data were collected with very well documented geological
descriptions and it was possible to determine that these results had been influenced
most strongly by their waste rock content. This suggests that when sampling for
rock characteristics, especially in breccia kimberlite facies, there may be a
requirement to carry out some form of indicator kriging to deal with the highly
skewed nature of the so-called ‘contaminating country rock fragments.’
Alternatively, a separate estimate of the proportion of dilution of each major type
might be used in generating the spatial model.

Comparison of estimation and simulation techniques for rock

characteristics

Estimation of individual rock property variables using linear estimation techniques,

such as ordinary kriging, will produce results that are unbiased provided a number
of assumptions are met. One of these requires that the variables of interest are
additive. The additivity property of a variable requires that the arithmetic mean of
a variable at an unsampled location can be calculated, without bias by a linear
weighted combination of the available data. There are some variables, including
many of those that describe rock characteristics response variables, that are not
strictly additive. The values at unsampled locations can technically be estimated at
sample scale using linear methods, but these estimates will be biased in some way.
This approach may in a few unique cases be acceptable when resulting bias in the
estimate may not have a material impact on the use of the biased estimates.
Unfortunately, it is not always possible to determine the magnitude and sign of the
bias nor to reliably identify when the resulting bias may have material impact.

As an example of a non-additive response variable, t10 represents the mass

proportion of a sample that is below a given size fraction following a controlled
breakage test. In some cases (e.g. if the samples were all of the same support and
density) it would be possible to transform this variable using data from the sample
to the mass of the sample that is below a given size for a given energy input, a 'mass
by mass' variable much like metal grades. In the post processing of these estimates
it would be possible to predict fragmentation and mass flows by using the energy
size relationship presented in Equation 13.

Carrasco, Chiles and Séguret (2008) present a mechanism to test the consequences
of treating a non-additive variable as additive. In this model a comparison is made

120 Stephen Coward - April 2020

 Chapter 5: Modelling physical characteristics of orebodies

between the 'illegitimate average of ratios' and the 'legitimate ratio of averages' as
presented in Equation 10.

𝒁𝒓
𝑹= Equation 10
𝒁𝒉

Using a similar notation, the variables in this case would represent :

𝒁𝒉 as the in situ sampled t10 can be treated as the in situ ratio, or proportion, and
can be incorrectly estimated using kriging from samples into blocks.

𝒁𝒓 can be considered the size distribution that results from crushing at the
treatment plant derived by correctly converting t10 into the mass proportion in a
block of a given size; and

𝑹 is the response variable the ratio of the treated and mined t10. represents the
process response variable of tonnes to undersize of 2mm

In this way it is possible to calculate the R variable in two different ways and then
determine the impact of the underlying bias that arises from averaging a non-
additive variable. :

The difference between the two approaches will be 'immaterial' when:

𝒁𝒉 is constant, or 𝒁𝒓 is a constant;

the ratio of 𝒁𝒓 to 𝒁𝒉 is a constant;

or the range of the values of 𝒁𝒓 , 𝒁𝒉 and the change in the ratio of 𝒁𝒓 to 𝒁𝒉 is very
small.

In the case here block size is assumed to be constant and so it is likely that the impact
of the bias of incorrectly using a linear geostatistical technique would be almost
immaterial. It is possible to krig and apply weights 𝑤𝑖 to the samples to estimate the
ratio of two variables ( 𝑍1 (𝑥) and 𝑍2 (𝑥) )where the data for both variables is
available at all sample Equation 11.

𝒏
𝒁∗𝟏 (𝒙) 𝒁𝟏 (𝒙)
∗ (𝒙) = ∑ 𝒘𝒊 Equation 11
𝒁𝟐 𝒁𝟐 (𝒙)
𝒊=𝟏

The ratio of the two estimated variables will be unbiased if ;

Stephen Coward - April 2020 121

 Evaluation of metallurgical recovery factors for diamonds recovered from kimberlites

• the weights sum to one ( ∑𝑛𝑖=1 𝑤𝑖 = 1 );

• the same variogram is used for the estimation of 𝑍1∗ (𝑥) and 𝑍2∗ (𝑥); and

• first order stationarity holds , at least within the search neighbourhood.

As a further caution, even if it is possible to mathematically justify an assumed

'quasi-additivity', the use of linear estimation techniques may still result in material
post processing errors, especially when the response of the process that will be
determined from the estimate involves physical blending of different rock types. In
the case of using t10 data to predict the fracture of mixtures of rock types may result
in material error as the energy transfer between rocks of different strengths is, in
many cases, non-linear and so the overall degree of fracture is related not only to
the t10 values but the differences between the t10 values that are being mixed.

It is also important to distinguish between estimating sample-size volumes and

estimating larger volumes. The variance of the latter must, of course, be less than
the variance of the sample data. The estimates of the larger volumes will also have
lower variance than the real sample volumes, but for larger volume data, we don’t
know the real variance.

Spatial simulation on the other hand will produce models of block values that have
a higher variability than estimated kriged models which may be useful for assessing
the impact of rock property variability on process performance. The assumptions of
the data properties for spatial simulation are more onerous and require sufficient
sampling to enable modelling of the histogram of the data and testing of geological
domains to ensure that assumptions of stationarity hold.

A demonstration of the differences between methods of estimating and simulating

the t10 for the study area is provided in this section. The approaches demonstrated
include:

• A test for additivity of the t10 variable;

• Independent kriging of the t10 values based only on the t10 data;

• Co-kriging of t10 with acoustic velocity and density;

• Sample support scale modelling the t10 at locations in each hole that have not
been sampled; and

• Simulation of t10 using a turning bands method.

122 Stephen Coward - April 2020

 Chapter 5: Modelling physical characteristics of orebodies

The suggested methodology begins with an up scaling of all the collected data to a
70cm composite. In this way the support of the destructive data (t10, and t10 sample
density) and the geophysical data (P wave slow) is equalised. Variograms were
modelled and the estimation parameters set (Table 14). Estimates were run on two
different search neighbourhoods, the second far larger than the first, as the first pass
estimation did not populate all blocks.

P-wave Velocity Sample Density t10_E1

Variable Nugget Range (m) Sill

C0 % of Sill Structue 1 Structure 2 Structure 1 % Structure 2 %
P_wave Velocity 0 0 2.93 6.48 54 46
Density 0.00046 17 2 6.72 45 38
Drop Weight t 10 3.98 33 1.1 6.1 39 28

Table 14:Table and graphics of the modelled variograms for P-wave Velocity, Specimen
Density and drop weight data.

Table 15 gives the summary statistics for the variables selected for estimating and
the statistics of the estimates generated into blocks of 0.7m x 0.7m x 0.7m

Stephen Coward - April 2020 123

 Evaluation of metallurgical recovery factors for diamonds recovered from kimberlites

VARIABLE Count Minimum Maximum Mean Std. Dev Variance CV

Pwave_Slow_Krig 427,004 162.52 308.89 249.36 19.71 388.49 1.56
Pwave_Slow_Mean 1,283 144.51 318.79 254.2 25.6 655.15 2.58
% Shift 12 -3 -2 -23 -41 -40

VARIABLE Count Minimum Maximum Mean Std. Dev Variance CV

T10_Density_Krig 376,094 2.43 2.85 2.54 0.04 0 0
T10_Density_Mean 114 2.43 2.85 2.54 0.07 0 0
% Shift 0 0 0 -43

VARIABLE Count Minimum Maximum Mean Std. Dev Variance CV

T10_E1_Krig 376,094 16.02 41.51 33.34 2.71 7.33 0.22
T10_E1_mean 114 16.02 41.51 33.3 4.04 16.29 0.49
% Shift 0 0 0 -33 -55 -55
Table 15: Summary Statistics for Data and first pass estimates.

In all cases the average of the raw variables and the estimates differ by less than 3%,
as expected the coefficient of variation of the kriged estimates is lower than that of
the raw data. The number of blocks populated for 'T10_density' and t10_e1 are far
fewer than that for P-wave. This is a result of the range used for P-wave variable
being far greater than that of the other two variables.

An analysis of the additivity of t10

Figure 37: Histograms of the original t10 values (left) and the transformed variable "mass
less 5mm in g/tonne" (Right).

The t10 transformation used is displayed in Equation 12. This transform converts the
proportion of the sample passing one tenth of the original rock size to a 'grade' like
variable which expresses the rocks amenability to crushing in the grams of rock that
would be below a certain size for a given energy input.

124 Stephen Coward - April 2020

 Chapter 5: Modelling physical characteristics of orebodies

𝑴𝒂𝒔𝒔𝑳𝒆𝒔𝒔𝑻𝒉𝒂𝒏𝟓𝒎𝒎 = (𝒕𝟏𝟎 ∗ 𝑫𝟏 ∗ 𝑽𝒔 ) × 𝟏𝟎𝟎 Equation 12

Where:

𝑡10 is the percentage by mass passing 1/10th of a particles original size for a given
energy input;

𝐷1 is the average density of the sample tested; and

𝑉𝑠 is the volume of the sample that was tested.

In this case the data comprise 116 values, at sample scale. The average of the
sampled t10 value is 33.29, which is the average of the ratios. To calculate the ratios
of the averages the average mass less than 5mm was divided by the average mass of
the samples to give a value of 33.262. The difference between the average of the
ratio and the ratio of the averages is 0.107% of the average of the original t10 values.

Ordinary kriging was carried out on both variables and the sample density to
populate the test block grid (as described in section 4.4). The block models showing
the variables estimated are shown in Figure 38. The average of the block estimates
for t10 was 33.54, and the average of the mass less 5mm was 8.44 g/tonne.

Figure 38: 3 dimensional projections of the kriging of drop weight values (left) and the
transformed variable "Mass less 5mm in g/tonne" (Right).

The result of dividing the average undersize by the average mass of the block (the
ratio of the averages) is 32.13, which is a difference of just over 5%. The magnitude
of the difference is sensitive to the change in the density estimate which in turn
affects the variation of the mass in each block.

Stephen Coward - April 2020 125

 Evaluation of metallurgical recovery factors for diamonds recovered from kimberlites

This data set shows a range of values that is relatively narrowly dispersed - with a
coefficient of variation for the t10 samples of 12% and density of 3%. At block scale
the coefficients of variation reduce to ~5% for t10 and a density coefficient of the
blocks of ~1.2%. This does not prove that the t10 variable is additive but gives some
indication that for this specific set of data the risk of treating t10 as an additive
variable may result in a 5% bias in the estimation.

Ordinary kriged estimates

Figure 39 shows cross-sections and histograms of the estimates for each variable
when they are up scaled to mining blocks sizes of 5m x 5m x 5m. It is noticeable that
there is now a considerably higher degree of smoothing.

Figure 39: Cross sections of the orebody and histograms for the three variables estimated
independently into a 0.7m x0.7m x 0.7m grid and accumulated into a 5m x 5m x 5m grid.

Simulation of rock characteristics

The same data were then used to generate a turning bands simulation of the same
area of the deposit. To do this the data variables were converted to Gaussian
variables using a Gaussian anamorphosis. The impact of the transformation is
depicted in Figure 40.

126 Stephen Coward - April 2020

 Chapter 5: Modelling physical characteristics of orebodies

Figure 40: Histograms showing transform of data from raw data to Gaussian variables.

These Gaussian data were used to develop the semi-variogram models depicted in
Figure 41. The three primary directions all seem to have similar ranges, with the
structure of the P-wave variograms being far more stable than those of the
variograms for the t10 data.

Figure 41: Variogram models (top left and bottom right)and cross variogram models
(bottom left) for the Gaussian transforms of P-wave and drop weight test data.

Stephen Coward - April 2020 127

 Evaluation of metallurgical recovery factors for diamonds recovered from kimberlites

These inputs were used to generate 50 spatial realisations of the variables at block
scale. The simulated values were back transformed, and the simulations were
validated by comparing the simulated histograms replication of the input data.
Perspective views of cross-sections of these variables comparing the texture of the
simulated values vs the kriged values are shown in Figure 42.

Figure 42: Perspective plots showing a cross-section through the test area for 5m x 5m x 5m
blocks for simulated and kriged p-wave velocity (upper images) and drop weight test
data(lower images).

The statistics for the block are given in Table 16 and show the differences in
averages, std. deviation and extreme values. As expected, the estimates have a lower
variance than the simulated values.

In this specific case, the differences between the simulated and estimated maximum
and minimum values are relatively small. This is to some degree a function of the
relatively closely spaced holes (5m) whereas in normal production models the drill
spacing will be far larger and hence the difference in variance between the
estimated and simulated values will be greater.

128 Stephen Coward - April 2020

 Chapter 5: Modelling physical characteristics of orebodies

Estimate Type Count Minimum Maximum Mean Variance

P-Wave Kriged 1600 194.34 287.07 248.05 351.95

P-Wave Simulated 1600 161.43 293.34 247.34 547.18

t10_e1_Kriged 1600 26.71 39.48 33.39 5.46

t10_e1_Simulated 1526 16.88 41.32 32.33 19.88

Table 16: Summary statistics comparing estimated and simulated values for P-wave velocity
and drop weight results at the 5m block scale.

Implications of different estimation and simulation pathways

The objective of this analysis is to demonstrate how different transformations of the

primary variables, and subsequent estimation of the transformed variables, may
impact on prediction of response variables at operational scale.

If it is assumed that the purpose of the operating strategy for a crushing and grinding
circuit is to achieve a given target grind size and allow throughput to vary, it is
possible to predict throughput in each period for a given input rock hardness. To
convert the measured in situ variable t10, to a prediction of throughput requires
several steps including:

• Measure the fracture response of a rock at three different energy

inputs;

• Model an energy breakage relationship, and capture the “A” and “b”
parameters of the fitted function;

• Use the product of the “A” and “b” parameters to predict the energy
consumption;

• Given an assumed installed comminution power predict the

throughput; and

• Assign a block mining sequence to evaluate the throughput achieved

in different periods of operation.

Stephen Coward - April 2020 129

 Evaluation of metallurgical recovery factors for diamonds recovered from kimberlites

The alternative pathways through these steps are depicted in Figure 43.

Figure 43: Schematic of the optional routes to use point scale sample data to predict
throughput.

The mechanics of each of the calculation steps are briefly described below:

Fracture response to energy consumption

The procedure for sample testing requires that several fragments of rock are subject
to a range of controlled, known input energies and the sizes of the resultant broken
particles are measured and expressed as the percentage of material passing one
tenth of the fragments’ original size (the t10) for a given energy input. The
relationship between increasing energy input expressed as kWh/t and the
increasing percentage of material passing one-tenth of its original size can be
plotted. A so-called breakage function can be fitted through these points. Figure 44
shows a plot for the case where fragments of the sample have been exposed to three
different energy levels, which have produced three different levels of fracture.

130 Stephen Coward - April 2020

 Chapter 5: Modelling physical characteristics of orebodies

Sample Comminution Function

35

30

T10 % passing one thenth of size

25

20

15

10

0
0 0.2 0.4 0.6 0.8 1 1.2
Specific Comminution Energy(Kwh/t)

VENODS1008 Predicted

Figure 44: A Relationship between input energy and degree of fracture, expressed as
percentage passing 1/10th of original particle size (blue diamonds), showing a fitted
breakage function in black.

Equation 13 describes The function that can be fitted though the data:

𝒕𝟏𝟎 = 𝑨(𝟏 − 𝒆−𝒃×𝑬𝒄 ) Equation 13

In this case the A value is 44 and the B value is 1.06, giving a A*b value of 47.65.

Using modelled breakage function parameters to estimate power consumption

Several authors (e.g., Bye et al., 2011) have used the product of the parameters fitted
to the breakage function (A*b) value to relate the t10 value directly to milling energy
consumption (although in reality the achieved throughput is a function of a host
more variables there it has been demonstrated that there is some relationship and
hence this could be a good variable to use to predict throughput for a target grind
size). This raises the issue of the impact of selecting a ‘controlling’ or ‘target
operational response variable’ for estimation and allowing others to fluctuate in
response to operation with the objective of meeting a targeted response.

This behaviour arises from the interaction of ore characteristics and an operational
strategy and not strictly only the properties of the ore being treated and hence could
be misleading if estimated into the block model without careful and precise
documentation of the assumptions and models used.

Stephen Coward - April 2020 131

 Evaluation of metallurgical recovery factors for diamonds recovered from kimberlites

Ideally in the value chain modelling, all response variables should be determined ‘at
run time’– e.g., the value chain model constraints would include a policy that the
process targets a grind size of p80 of 75 µm and allows the feed rate to fluctuate. A
separate run of the value chain model might alternatively target 1000tph feed rate
and let the grind size oscillate in full response to the input rock properties. In both
cases the relationship between the primary variables and response variables will be
very different, so the primary variables should be in the block model and the
response and operating strategy should be calculated in the value chain model.

Figure 45: Plot showing broad correlation between average long run energy consumption
in Semi Autogenous Grinding (SAG) mills and average orebody A*b values (Daniel, Lane and
McLean, 2010).

Mapping Energy Consumption to Throughput

In this model, a simple inverse relationship is used where increased energy

consumption of the rock breakage reduces throughput; this relationship is depicted
in Figure 46 . This is based on prior performance data observed at the sample site.

132 Stephen Coward - April 2020

 Chapter 5: Modelling physical characteristics of orebodies

Power Consumption vs Throughput factor

1.2

Predicted Throughput Factor

0.8

0.6

0.4

0.2

0
0 2 4 6 8 10 12 14 16 18 20
Estimated Power Consuption in Block (Kwh/t)

Figure 46: Relationship between power consumption and throughput for a target grind
size.

Mine planning and block sequencing

The mine plan was simplified to a layer-by-layer plan to assign each block a mining
sequence – the blocks then mined could be assigned a time to crush based on their
t10, which can then be cumulated into a daily, weekly and monthly throughput.

Several pathways are possible to convert the sample data into block estimates of
throughput assuming that the process will be constrained to deliver a fixed size
distribution. These options are briefly described in Table 17.

Stephen Coward - April 2020 133

 Evaluation of metallurgical recovery factors for diamonds recovered from kimberlites

Pre- Block
Spatial Post estimation
# Point data derivation estimation scale Comments
estimation calculation
calculation prediction

In each block use t10’s are a proportion

Break sample sieve and
three t10 to fit curve by mass of the sample
1 determine % passing 1/10th of
original particle size, for three
Co simulate
or estimate
and derive A and b passing a given size. If
None value, Multiply A by Throughput the blocks are of the
energy levels. Generates three- three t10
b- then used A*b same size and density
point values at each sampled values
through function to t10 can be considered
X, Y Z location
predict throughput additive

As A’s and B’s are

Fit function to
2 Break sample sieve and
determine % passing 1/10th of
three points at
Estimate A
Use A*B in each
block, put through a
parameters of curves
tests for additivity and
each sample Throughput non-linearity would be
original particle size, for three and B values transfer function and
location, derive A required
energy levels derive throughput
and B value

Fit function to
three points at
each sample
location, derive A Spatially
3 Break sample, sieve and
determine % passing 1/10th of
and B value, estimate and
Calculate energy
consumption and
calculate A*b and or simulate Throughput
original particle size, for three throughput in the
estimate A*b, product of
energy levels. block
Calculate energy A*b
consumption and
throughput in the
block

Fit function to
three points at
4 Break sample, sieve and
determine % passing 1/10th of
each sample
location, derive A
Estimate
Calculate throughput
Energy Throughput
original particle size, for three and B value, in the block
Consumption
energy levels. calculate A*b,
Calculate energy
consumption

Fit function to
three points at As throughput is a rate
each sample Spatially variable is subject to
5 Break sample, sieve and
determine % passing 1/10th of
location, derive A
and B value, put
estimate and
or simulate
constraints and limited
to some maximum and
None Throughput
original particle size, for three A*B through the minimum. This
energy levels. transfer function, throughput approach may estimate
Calculate variable infeasible throughput
throughput at values
point

Use correlation
Break sample, sieve and In each block use
between t10s and
determine % passing 1/10th of three t10s to fit curve
6 original particle size, for three
Geophysical
response to
Co simulate
or estimate
and derive A and b t10 are a proportion by
energy levels, where there are values, Multiply A*B- Throughput mass of the block, thus
calculate t10’s at three t10
no samples down the core then use A*B through variable is additive
un-sampled values
used Geophysical relation to function to predict
locations down
t10 to augment core values. throughput
core

Run value chain

Use correlation model that
Break sample sieve and
between t10s and incorporates a
determine % passing 1/10th of
7 original particle size, for three
Geophysical
response to
Co simulate
or estimate
process simulator,
treats groups of
Incorporates inter-
block interaction and
energy levels, where there are Throughput
calculate t10’s at three t10 blocks, dynamic dynamic effects of
no samples down the core
un-sampled values population balance operation
used Geophysical relation to
locations down models, predicts
T10s to augment core values
core many aspect of
process response

Table 17: Options for use of spatial data to estimate process response variables.

Outcomes from pathway one

In this pathway, the individual laboratory results achieved from crushing each
sample are converted to three t10 values treated as point data. The three t10 values
are spatially estimated using ordinary kriging.

134 Stephen Coward - April 2020

 Chapter 5: Modelling physical characteristics of orebodies

Strictly speaking t10 is not an additive variable. It could however be transformed to

a mass proportion variable and then kriged. It would be possible to do a back
transform to t10 by recalculating, in each block, the proportion of the total mass of
material that is less than 1/10th of its original size. In this synthetic example all of
the blocks in the model have the same volume and density and thus the transform
to a mass variable is not required. (See tests for Additivity in Section 5.5)

The resulting values of t10 in each block were analysed and a least squares curve was
fitted. The A*b value for each of these curves is used deterministically to calculate
power consumption and throughput. Summary statistics for the block estimates are
shown in Table 18.

Proportion of Time to
Predicted
max treatblock Days
t10_e1 t10_e2 t10_e3 Model A Model b A*b Power
throughput (Portion of Treated
Consumption
possible day)
Minimum 7.58 16.63 26.71 35.45 0.06 34.98 5.46 0.65 0.31 0.31
Average 10.92 20.92 33.39 60.05 0.92 51.17 7.50 0.96 0.33 266.58
Maximum 13.79 24.01 39.48 562.93 1.75 65.84 11.65 1.00 0.48 522.60
Std. Deviation 1.15 1.35 2.34 23.70 0.24 5.35 0.99 0.07 0.03 150.54
Cof. Var. % 10.53 6.47 7.00 39.47 25.75 10.45 13.20 6.80 7.94 56.47

Table 18: Summary statistics for the estimation and calculation of block scale properties
for pathway 1.

Pathway two

In this pathway, the t10 values at the point are plotted as described in Figure 44 and
the A and b parameters fitted. The estimation process estimates the A and b values.
Once the estimate is completed the A and b values in each block are multiplied
together to generate the A*b product value which is used to calculate the resulting
energy consumption and throughput (Table )

Proportion Time to
Predicted
of max treatblock Days
t10_e1 t10_e2 t10_e3 Model A Model b A*b Power
throughput (Portion of Treated
Consumption
possible day)
Minimum 7.58 16.63 26.71 44.53 0.43 42.64 2.64 0.83 0.31 0.31
Average 10.92 20.92 33.39 65.28 0.96 60.51 6.26 0.99 0.32 254.37
Maximum 13.79 24.01 39.48 250.46 2.05 120.73 9.19 1.00 0.38 505.09
Std. Deviation 1.15 1.35 2.34 20.20 0.18 10.95 1.16 0.03 0.01 145.46
Cof. Var. % 10.53 6.47 7.00 30.94 18.51 18.10 18.55 2.78 3.08 57.18

Table 19: Summary statistics for the estimation and calculation of block scale properties
for pathway 2

Pathway three

In this pathway, the t10 data are used to derive a breakage function, the A and b
values are multiplied, and this product is estimated. The statistics are summarised
in Table 20.

Stephen Coward - April 2020 135

 Evaluation of metallurgical recovery factors for diamonds recovered from kimberlites

Proportion Time to
Predicted
of max treatblock Days
t10_e1 t10_e2 t10_e3 Model A Model b A*b Power
throughput (Portion of Treated
Consumption
possible day)
Minimum 7.58 16.63 26.71 44.53 0.43 38.37 5.48 0.73 0.31 0.31
Average 10.92 20.92 33.39 65.28 0.96 51.93 7.36 0.97 0.32 263.96
Maximum 13.79 24.01 39.48 250.46 2.05 65.57 10.43 1.00 0.43 519.13
Std. Deviation 1.15 1.35 2.34 20.20 0.18 5.37 0.95 0.06 0.02 149.94
Cof. Var. % 10.53 6.47 7.00 30.94 18.51 10.34 12.84 5.96 6.68 56.81

Table 20: Summary statistics for the estimation and calculation of block scale properties for
pathway 3.

Pathway four

This pathway generates estimates of power consumption values that are then
converted to throughput by calculation. The summary statistics for the outputs of
this process are given in Table 21.

Proportion Time to
Predicted
of max treatblock Days
t10_e1 t10_e2 t10_e3 Model A Model b A*b Power
throughput (Portion of Treated
Consumption
possible day)
Minimum 7.58 16.63 26.71 44.53 0.43 38.37 5.52 0.57 0.31 0.31
Average 10.92 20.92 33.39 65.28 0.96 51.93 7.57 0.95 0.33 269.98
Maximum 13.79 24.01 39.48 250.46 2.05 65.57 13.34 1.00 0.55 527.07
Std. Deviation 1.15 1.35 2.34 20.20 0.18 5.37 1.04 0.07 0.03 152.22
Cof. Var. % 10.53 6.47 7.00 30.94 18.51 10.34 13.69 7.32 8.45 56.38

Table 21: Summary statistics for the estimation and calculation of block scale properties
for pathway 4.

Pathway five

In this pathway, the t10 data are used to derive A*b values and these are used to
calculate the energy consumption and the throughput. The point scale calculated
throughput values are then used to generate estimates of throughput at block scale.
The statistics of the output are given in Table 22.

Proportion Time to
Predicted
of max treatblock Days
t10_e1 t 10_e2 t10_e3 Model A Model b Axb Power
throughput (Portion of Treated
Consumption
possible day)
Minimum 7.58 16.63 26.71 44.53 0.43 38.37 5.52 0.71 0.31 0.34
Average 10.92 20.92 33.39 65.28 0.96 51.93 7.57 0.93 0.34 274.86
Maximum 13.79 24.01 39.48 250.46 2.05 65.57 13.34 1.00 0.44 538.14
Std. Deviation 1.15 1.35 2.34 20.20 0.18 5.37 1.04 0.06 0.02 155.66
Cof. Var. % 10.53 6.47 7.00 30.94 18.51 10.34 13.69 6.08 6.53 56.63

Table 22: Summary statistics for the estimation and calculation of block scale properties
for pathway 5.

Although pathways 5 and 6 have been given in table 2 above the outputs of these
pathways will be discussed in Chapter 6 as they require a dynamic process model to
evaluate these approaches to estimation pathways.

136 Stephen Coward - April 2020

 Chapter 5: Modelling physical characteristics of orebodies

Comparison of outcomes of different methods

The distribution data for each of the variables calculated and/or estimated are
shown in Figure 47.

Figure 47: Histograms of the variables that are calculated and estimated in pathway 1.

In these histograms, the data have been grouped by the resulting throughput factor.
This figure clearly shows the relationship between the variables used to derive the
throughput indicator. These distributions can be contrasted with the histograms for
pathway 3 depicted in Figure 48.

Figure 48: Histograms of the variables that are calculated and estimated in pathway 3, data
has been grouped by predicted throughput quartiles.

In pathway three the estimation of the A*b variable results in a distribution of the
A*b variable that is less dispersed than that for pathway 1. This ultimately results in
a far higher estimate of the throughput factor.

Stephen Coward - April 2020 137

 Evaluation of metallurgical recovery factors for diamonds recovered from kimberlites

Comparison Weekly Throughput

Prediction
6500
6300
Weekly Throughput (tonnes)

6100 T10
t10 Estimated
Estimated
5900
A and b values Estimated
5700
5500 A*b Estimated
5300
Power Consumption
5100 Estimated
4900 Throughput Estimated
4700
4500
0 10 20 30 40 50

Figure 49: Plot showing the variability in weekly throughput and a table with summary
statistics for each of the variables calculated through each pathway.

The implications of this can be demonstrated by reviewing the different predictions

that result from using the different approaches. Figure 49 shows how the average
values differ, but perhaps more importantly also shows how the weekly predicted
throughout would vary. In some weeks the difference exceeds 22% of the mean
throughput prediction, and on a block-by-block scale this would be even more
significant.

5.6 Generic approach and considerations for generation of

suitable spatial models

Figure 28 depicts a pathway to generate a spatial model that depends on the sample
data available, the variable being considered and the use of the estimate and ore

The spatial modelling of rock properties has potential to contribute substantially to

the value of mining projects. To do this successfully requires integration of tools and
techniques from several fields, including geology, mining, metallurgy and
Geostatistics. There are several ways in which the modelling processes can be
improved.

Improved data analysis at existing operations, which includes retrospective analysis

of process response to orebody characteristics at an operational scale, will provide

138 Stephen Coward - April 2020

 Chapter 5: Modelling physical characteristics of orebodies

an opportunity to develop a generic understanding of the important orebody

properties that impact on different unit responses and that have systematic impacts
on operational performance.

Development of an iterative and integrated approach to sampling, measurement

and variable estimation is an important strategy to follow when trying to design and
execute rock characteristic estimation programmes.

Geophysical tools can yield data that, although weakly correlated with destructive
tests, can achieve a higher spatial coverage at a far lower cost than destructive tests.
Using this additional data has potential to improve the estimates of spatial rock
characteristics in the following ways:

• Identification of potential bias in sample layouts used for destructive testing;

• Including this weakly correlated data in the rock characteristic estimation

process; and

• Identifying anomalous results.

The effective use of geophysical tools however requires that careful attention is
given to the accurate location of the geophysical data and careful and repeated on-
site calibration of the sondes used. Even a small deviation of the order of 0.5m can
destroy the correlation required to make this data useful. As demonstrated here the
acquisition of geophysical data from both the core and the hole that it is extracted
from can be used to correct core location data and assist in determining the
relationship between the minerals and geology of the rock that is driving the
measured geophysical response.

The collection of destructive data acquired from small rock specimens requires that
specific attention be paid to the execution of the test themselves. As it is impossible
to crush the same rock twice, the quality control measures are centred on
repeatability of the testing procedure. Indications are that the introduction of
geophysical measurements of core just before it is destructively tested would
improve laboratory quality control. This would also help to develop understanding
of the relationship, at core scale, between the geophysical responses and the
measured destructive response.

Several multivariate techniques can be used to identify relationships between the

properties of the rock and the responses that are measured. PCA and PLS regression
models demonstrated here, are used to augment the destructive data with
geophysical responses. Acoustic velocity measurements show a correlation with the

Stephen Coward - April 2020 139

 Evaluation of metallurgical recovery factors for diamonds recovered from kimberlites

elastic moduli of the rock. These in turn have been shown to be correlated with the
destructive t10 data.

The massive differences in the scale of laboratory test work and operations require
that existing approaches to up-scaling be carefully evaluated. The impact of the size
of specimens used in destructive rock property tests on the data dispersion and
variogram range requires further investigation. Larger specimens may reduce the
error in block scale estimates of rock property.

Response variables, such as throughput, can be calculated and estimated in different

ways. It has been demonstrated that the choice of the pathway has an impact on the
resulting block scale estimate. It is recommended that where possible the most
primary version of the data that are gathered at small scale be used in the estimation
process.

In the next chapter, the use of spatial rock characteristic estimates in process
simulations will be explored.

140 Stephen Coward - April 2020

 Chapter 6: Process simulation and rock characteristics

6 PROCESS SIMULATION AND

ROCK CHARACTERISTICS

6.1 Introduction

The methodology suggested here to evaluate the metallurgical recovery factor

requires a ‘value chain simulation’ that comprises multiple realisations of the
orebody and a framework that can be used to evaluate the impact of variable
kimberlite characteristics on diamond recovery. This processing framework
requires unit process models that can be used to simulate the mining and treatment
processes that respond to the variable and uncertain kimberlite characteristics.

The chapter begins with a review of open pit mine models and a discussion of their
limits and constraints is discussed. A method to model and simulate the process of
ore extraction is provided. A description of the outputs of the mining model, and its
use in the unit process models in the treatment plant simulation is given.

The remainder of the chapter describes a methodology for formulating an

integrated model that can be used to evaluate how uncertain and variable rock
properties impact on the performance of the treatment process. The application of
the model is described in the subsequent chapters, where the integrated model is
used and demonstrated in two case studies that analyse a number of operational
configurations, assess diamond recovery uncertainty and improve the definition of
the range of the expected metallurgical recovery factor.

Three, increasingly complex, diamond recovery models can be used to estimate the
metallurgical recovery factor. The approaches are depicted in Figure 50 and
include:

a) A model based on the extension of micro and macro diamond size

distributions to determine total diamond content and size distribution;
b) A model based on the relationship between total kimberlite grind, diamond
liberation and lockup; and
c) A model based on a plant process simulation that incorporates a calibrated
processing model of the diamond recovery plant.

Stephen Coward - April 2020 141

 Evaluation of metallurgical recovery factors for diamonds recovered from kimberlites

The micro-macro diamond model approach is suited to early stage projects where
little is known about the deposit. In such a model only a few macro diamond samples
are available but the relatively large number of micro diamonds (of the order of
100’s of stones per kg) that occur in small samples of diamond bearing ore can be
used to model the in situ diamond distribution (Chapman and Boxer, 2003; Caers
and Rombouts, 1996). This modelled in situ size distribution can be used to predict
the abundance of coarse diamonds and in some cases the expected macro diamond
recovery. This requires the use of relatively broad assumptions about mining and
treatment efficiencies to predict the expected diamond recovery at a specified top
and bottom size cut off.

The granulometry based liberation and lock-up model is suited to projects that have
a well-conceived overall process design and treatment flowsheet, but where there
is limited information on unit process efficiency and the impact of mining methods
or the ore properties on process rate, efficiency and cost. Several parameters of this
model are usually inferred from other operations. The ‘granulometry liberation loss
model’ relates the size distribution of the recovered diamonds to the discarded
kimberlite particles to estimate the ‘locked’ diamonds; i.e., those that are not
sufficiently freed from the host ore and are unrecovered or lost. The model provides
sufficient insight to begin trade-off studies between different flowsheet options.

A more detailed approach that uses calibrated population balance models can only
be taken when sufficient inputs are available to calibrate accurately the unit process
models. The population balance models can be adapted to include several unit
processes that are all impacted by the variable and uncertain properties of the
treated rock.

In most diamond operations a mass balance model that has several unit processes,
including at least a comminution circuit, a dense media separation circuit and a final
recovery plant, is required. With this model it is possible to estimate how the
changes in the feed characteristics impact on the recovered diamond grade and size
distribution.

142 Stephen Coward - April 2020

 Chapter 6: Process simulation and rock characteristics

Early stage evaluation

Micro diamond
Pre-Feasibility,
Model and
option evaluation
Global geometry
Feasibility and operation

Diamond size model

Granulometry model Diamond size distribution

model

Granulometry model

Process simulation

Figure 50: A schematic depiction of approaches to estimating the recovery factor for
different project maturities.

Approaches used to determine block sequencing and characteristic

alteration during mining

The design and selection of a mining process determines the sequence in which the
ore arrives at the process plant and the extraction methods that are used to fracture
the ore and transport it either directly to the plant, or to a stockpile or to dump as
waste. Design and optimisation of the ultimate pit, the mining sequence execution
process methods used to plan the schedule and sequence of blocks to be mined, and
model the impact that drilling, blasting and haulage have on the physical rock
characteristics (Dowd, 1976; Lane, 1998)

These usually aim first to determine an ultimate pit shell which contains the ore that
is profitable given the economic constraints of the project (Lerchs Grossman, 1965)
and expected contents and value of each mining block. This ultimate maximum size
excavation is then ‘optimised’ in terms of the sequence of push-backs or cuts to
excavate the ultimate pit. The schedule is then derived by determining the fleet size
and targeted treatment tonnage over relatively long periods. The result of the mine

Stephen Coward - April 2020 143

 Evaluation of metallurgical recovery factors for diamonds recovered from kimberlites

design work is to assign a sequence to each of the selective mining units (SMU) in
the block model. This is discussed in more detail later in this chapter.

Approaches to liberation modelling in the minerals industry

The modelling of crushing processes is required to predict comminution outcomes

(the size distribution of the rock product, the throughput rate, diamonds liberated,
and potential for diamond damage) for several feed types and process
configurations.

Initial models considered, that have been developed in the mining industry, require
the following assumptions for their predictions to be valid:

• Limited or very little mineral grain fracture;

• Breakage rates and selection criteria are independent of rock texture;
• Liberation is related to mineral grain size; and
• Liberation and comminution can be considered to be independent
processes.

Several liberation models have been developed, and according to Munn et al.
(Napier-Munn et al, 1999) two primary approaches can be distinguished.

The first approach is to evaluate the rock texture and then to develop predictive
models based on a mathematical description of the texture to relate the change in
the size of the rock to the change in liberation. These models include work by Gaudin
(1939), Weigel and Li (1967) and King (Beniscelli et al., 2000). They provide an
estimation of the possibility that fractures will intersect a diamond for a given stone
density and fracture size.

The second approach is to gather feed and progeny data from a crushing process
and analyse the degree to which valuable minerals have been liberated. Weedon
(1992) carried out several tests on crushed rock particles and reviewed the degree
to which the mineral of interest had been liberated. This data was then used to
establish characteristic curves that described the relationship between final rock
size and the expected liberation. Morell et al. (Box and Draper, 1987) demonstrated
that the liberation was largely independent of the process used to achieve
comminution. As a result, liberation could be predicted by a measure of the size
distribution of the contained mineral and the size distribution for the crushed ore.
Using this finding they could model and predict the liberation of sphalerite using a
three stream, two process population balance model. The three streams that were
modelled were the gangue, the locked valuable component and liberated valuable
component. The gangue and the liberated valuable component are processed

144 Stephen Coward - April 2020

 Chapter 6: Process simulation and rock characteristics

through conventional population balance model crushers (Box and Draper, 1987).
The locked valuable component stream is also processed through a conventional
crusher model and its products, locked and liberated mineral streams and gangue,
are added to the other products emerging from the other crusher models. Gay
(2002) has also produced similar models that reconstruct the parent composition
from the progeny that is observed in both valuable and discard streams.

In the sampling of diamonds, the micro-diamond recovery process recovers most of

the diamonds in a sample by dissolving the host rock. This suggests that very high
recoveries can be obtained and that the size distribution of the diamonds is not
materially changed during the recovery process. This contrasts markedly with the
bulk sampling of kimberlite for diamonds where the recovered, or observed,
distribution is compromised by the impact of losses in the process of sample
treatment. These losses include diamond damage, destruction of diamonds above
the top process size, loss of diamonds below the bottom cut-off size and lock up of
diamonds in the discarded coarse kimberlite. The micro-diamond distribution,
however, under-represents the coarse diamonds and can be viewed as giving a
truncated distribution. Despite this, the recovered micro diamond distribution gives
an indication of the size distribution of the valuable coarser component of in situ
diamonds. It is possible to establish a relationship between these micro and macro
distributions to model the liberation that is likely for a given size distribution of
processed kimberlite. This in situ size distribution is the main ‘texture’ feature that
is required for modelling liberation in the context of estimating the metallurgical
recovery factor.

Approaches to modelling diamond recovery at different project stages

The objective of modelling the treatment processes is to create a relationship

between:

• the estimated kimberlite properties;

• the contained diamond distribution; and
• the expected recovery efficiency measured in grade recovery and stone size
distribution recovery.

This model can be used to determine the quantity and size distribution of diamonds
that will be recovered from each estimated unit of ore in each period of mining.

As diamond projects move from prefeasibility to the feasibility study stage more
sampling of the orebody is carried out, to provide additional information on the
diamond grade, size distribution and value as well as the spatial nature of waste and
kimberlite characteristics. This information typically includes several

Stephen Coward - April 2020 145

 Evaluation of metallurgical recovery factors for diamonds recovered from kimberlites

measurements of the rock characteristics (measures of strength, fracture

properties, density etc) that facilitate the use of models to select critical items of
equipment, e.g. crushers, and estimate the relationship between the rock properties
and the unit expected average performance. The population balance modelling
approach is widely used in the industry for the design and configuration of process
plants and can be adapted specifically to estimate diamond recovery plant
efficiencies. In this more detailed approach to the modelling, the relationship
between diamond size, rock density and fracture properties becomes a vital
component in both estimating liberation and tracking the degree of diamond release
through the process.

The unit process models for comminution and DMS separation are data dependant
and require the fitting of model parameters are consistent with the observed data.
The fitting process is iterative and is assumed to be valid when the errors between
the observed data and the model are minimised. In brownfield situations this can be
time consuming but is generally achievable with well-considered and executed
process surveys. In greenfield situations it is necessary to rely on several diverse
sources of data to estimate process model parameters and to create a feasible
process simulation (i.e. one that converges to a stable solution.).

The primary focus of this research is to model the impact of variable and uncertain
rock strength on comminution, liberation and recovery of diamonds. There is a tacit
assumption that the equipment is well maintained and effectively operated. It is
however possible that operational personnel can at times have an influence on
process performance by adjusting process parameters (e.g., crusher gap setting).
Changes to the operating parameters of crushers, screens and dense media
separation units can either increase or decrease recoveries but will increase the
variability of the process output (Demming, 1986). These impacts have not been
explicitly considered in this research work, though some of these impacts may
contribute to the overall variance of the data that are obtained from operating mines
and is an area for further research.

6.2 Modelling of mining processes

Mining processes include drilling, blasting, material handling mixing and

transportation of the ore to the processing plant. The main implications of mining
processes that are considered in this research are:

• The impact of mining on the sequence and rate of mining; and

• The impact that the mining processes have on rock properties that are
considered to influence diamond recovery.

146 Stephen Coward - April 2020

 Chapter 6: Process simulation and rock characteristics

Usually, in the case of open pit mining of kimberlite pipes, all the kimberlite material
is mined and so the mined excavation extends to the boundary of the kimberlite .The
decision to leave sub-economic material in the pit is usually constrained by the low
structural strength of kimberlite and the need to ensure access to payable
kimberlite. In some cases where the pit is large enough (e.g. Williamsons mine)
areas of sub-economic kimberlite can be mined around, although this is not
common. In mines that contain relatively small un-payable material volumes (e.g.
Ellendale) a variety of ore/waste selection and rejection (grade control) practices
are used to determine the ideal destination for mined material i.e. waste or sub-
grade or plant feed. Given the difficulty and cost of sampling for grade these
practices are often based on a combination of the assigned grade in the
resource/reserve model, a visual assessment of the kimberlite as it is exposed, and
ease with which the location of the pipe contact with waste rock can be defined.

In the ‘Integrated Evaluation Model’ a method to replicate the block selection and
destination assignment has been developed. Each block in the grade model is
assigned a planned sequence number that is derived from the long-term mine plan
that is derived using conventional pit optimisation algorithms. The resulting
schedule is derived from the integrated model that responds to the interaction of
the orebody properties and the constraints of the mining method.

The mining process will impact on the rock strength properties, requiring some
modifier to the primary properties that have been spatially estimated or simulated
into the block model. The relationship between blasting parameters and diamond
distribution has been investigated by Guest et al. who found that within four blast
hole diameters most of the contained diamonds would be destroyed. (Guest, 1997)

Impacts of drilling and blasting on ore characteristics

To extract material from the mined pit, many operations use a drilling and blasting
method. Increasing the efficiency and reducing the cost of this operation has been
the focus of considerable research (Wilmott 2004, McGee 1995) The blasting will
produce a range of sizes of rock and the impact that this has on the process will
include the total power required as well as the mass flow in the primary section of
the comminution circuit. The impact on total grind will, however, primarily be
related to the total energy used in the blast, commonly referred to as the powder
factor measured in kg of explosives used per tonne of ore mined.

It is also important to consider the residence time of the ore post blasting in the pit
or on stockpiles where it is exposed to the atmosphere and will degrade to some
extent. In Kimberley in the late 1800’s this process of natural weathering was used
to limit the energy required to crush the kimberlite. At some sites the total content

Stephen Coward - April 2020 147

 Evaluation of metallurgical recovery factors for diamonds recovered from kimberlites

of clay and the clay species can be used to build a relationship between proportion
of the feed that has been exposed to rain and the overall comminution achieved

Materials handling stockpiling and storage

There is invariably some form of stockpiling on diamond mines, ranging from small
surge capacity bins (e.g., Snap Lake) to large +100,000 tonne stacker reclaiming
systems (Premier Mine). The impacts of blending and weathering need to be
correctly considered in any model of the operation as this process can, if correctly
operated, reduce the variability of grade and other important ore characteristics
such as waste rock content, UCS etc. (Robinson, 2004; Everett, 2001) This can be
relatively simplistically modelled in extreme cases by either not adapting block
characteristics which implies no blending or, for each characteristic of interest, the
entire load in the stockpile is averaged at each time increment in the simulation.
More realistic approaches include creating sub-parcels of several blocks that are
deemed to be co-located and the degree of mixing is controlled by some function of
both residence time and sequence of loading and withdrawal.

6.3 Modelling of diamond treatment plant process

Integration and adaption of population balance process models with an ore stream
that has multivariate characteristics is required to estimate the recovery factor.
Particular attention is given to the derivation of the parameters of the unit process
models, and how these are influenced by rock properties, and how they can be
perturbed realistically to quantify the range and variability of the process response
to changing rock properties.

Population balance modelling is used extensively in the minerals processing

industry and can be carried out on several software platforms but is typically used
on relatively advanced projects as the unit process models require calibration of
their parameters that are usually best derived from operating units that are
operating on the ore in question. In the absence of full-scale operational data, it is
however, possible to use small-scale pilot studies to derive the relationships
between ore characteristics and process responses.

Population balance models work by creating several ‘bins’ or “intervals’ for each
rock characteristic (e.g. size intervals, density intervals, grade intervals etc) that are
carried through the flowsheet and the masses entering and leaving each interval at
each stage of the flowsheet are required to balance. The balancing is achieved
through a set of iterative calculations and mathematical models that are required to
converge to some pre-set error tolerance. Using this approach for diamond
processing it is possible to model both the ore size distribution, the ore density

148 Stephen Coward - April 2020

 Chapter 6: Process simulation and rock characteristics

distribution as well as the diamond size distribution and to carry these through the
flowsheet.

A survey of De Beers’ operations provides an insight into the ranges associated with
the parameters that can be input into unit processes. The rock properties used
include the expected head feed size distribution, the strength of the rock expressed
by an energy size relationship using sampled t10 and a fitted breakage function, and
a distribution of density per size class.

Theoretical underpinnings of population balance models

Population balance model methodology

The aim of this approach is to simplify the quantitative description of particle

characteristics yet retain sufficient detail to enable process modelling and
simulation. A useful distinction can be made between internal and external particle
coordinates and distribution densities (King, 2001). The properties that describe
the individual particle properties are considered ‘internal coordinates of the
particle phase space,’ and the external coordinates are those that describe ‘location’
of the particle in the process in terms of the energy, stress, strain etc. to which the
particle will be subjected.

The internal coordinates should be sufficient to describe all significant

characteristics of the particle. These may include some characteristics that are
primary and some that are response. Even when considering primary properties
(e.g., particle density) these may have to be inferred or derived from the mineral
composition of the material that has entered a specific unit.

The external coordinates and their changes impact on the internal coordinates. This
relationship can be modelled in several ways and sets up a formalised framework
for developing and testing simulation of mineral processing flow sheets.

Notation conventions used by King (2001) to describe population model method

include:

𝜓(𝑥)𝑑𝑥 is the number fraction of particles per unit volume of phase space;

𝑥 is the coordinate point in phase space 𝑥𝜀ℝ𝑁 ;

𝜇 is a vector of velocities at which particles change their phase space;

Stephen Coward - April 2020 149

 Evaluation of metallurgical recovery factors for diamonds recovered from kimberlites

ℛ(𝑥) is the rate at which particles at coordinate position 𝑥 are destroyed, this rate
is specified as mass per unit volume of phase space per unit time;

𝑊𝑖𝑛 is the mass rate at which solid particles enter the system;

𝑊𝑖𝑛 is the mass rate at which solid particles leave the system;

̅ is the average mass of a particle at point 𝑥 in the phase space;

𝑚

𝑁 is the total number of particles in the system;

𝐴, 𝐵, 𝐷 are the rates of addition, birth and destruction of particles; and

𝑄 is the rate of removal of material through product streams.

𝐵(𝑥; 𝑥 ′ ) is a distribution function that describes how particles move discrete

distances in the internal phase space following breakage. It is the mass fraction of
material that has phase co-ordinate less than 𝑥 after a destructive event of particles
originally located at 𝑥′. This suggests that every phase co-ordinate of the particle is
now less than the corresponding element of 𝑥. The associated density function for
breakage is defined as b(𝑥; 𝑥 ′ ) and

a(𝑥; 𝑥 ′ ) is the distribution density function for particles produced by attrition and
wear of particles at 𝑥′

It is possible, as described by King (2001) to define a region 𝑅𝑐 of the particle phase

space enclosed by a surface 𝑆𝑐 and then develop a model that accounts for all
particles arriving and leaving this area.

𝝏
∫ 𝑵𝝍(𝒙)𝒅𝒙
𝝏𝒕 𝑹𝒄
Equation 14
= − ∫ 𝑵𝝍(𝒙)𝒖. 𝒏𝒅𝝈 − 𝑫 + 𝑩 − 𝑸
𝑺𝒄
+𝑨

where 𝑵 is the ‘outward pointing’ normal vector to the phase space 𝑆𝑐 at point 𝑥.

The destruction, birth, material removal and arrival terms are described below.

150 Stephen Coward - April 2020

 Chapter 6: Process simulation and rock characteristics

The destruction processes are represented by:

𝓡(𝝍(𝒙), 𝒙, 𝑭[𝝍(𝒙)]
𝑫=∫ 𝒅𝒙 Equation 15
𝑹𝒄 ̅ (𝒙)
𝒎

Where D is the number of particles broken per unit time in the control volume 𝑅𝑐 .
The rate of breakage is a function of the number distribution of both the particles in
the control volume and of the entire distribution function.

The birth processes are given by:

𝑩
𝟏
=∫ ∫ 𝓡(𝝍(𝒙′ ), 𝒙′ , 𝑭[𝝍(𝒙)])𝐛(𝒙; 𝒙′ )𝒅𝒙′𝒅𝒙
̅ (𝒙) 𝑹′(𝒙)
𝒎
𝑹𝒄 Equation 16
𝑵
−∫ ∫ 𝝍(𝒙′ ), 𝝁(𝒙′ ), 𝑭[𝝍(𝒙)])𝐚(𝒙; 𝒙′ )𝒅𝒙′𝒅𝒙
𝑹𝒄 ̅
𝒎 (𝒙) 𝑹′′(𝒙)

In Equation 16 𝑅′(𝑥) and 𝑅′′(𝑥) are the regions of phase space where particles that
have changed their size either by breakage or attrition and can enter the phase
volume 𝑑𝑥 around the point 𝑥.

The addition of material in this phase space can be described as:

𝑵
𝑨 = 𝑾𝒊𝒏 ∫ ( ) 𝝍𝒊𝒏 (𝒙)𝒅𝒙 Equation 17
𝑹𝒄 𝑴

And the removal rate as:

𝑵
𝑸 = ∑ 𝑾𝒐𝒖𝒕 𝒋 ∫ ( ) 𝝍𝒐𝒖𝒕 𝒋 (𝒙)𝒅𝒙 Equation 18
𝒋 𝑹𝒄 𝑴

The above equations can be integrated over the surface of the reference region. King
(2001) however suggests that the application of the divergence theorem, which
equates the integral of the phase space volume 𝑅𝑐 .with that of the enclosing space
surface 𝑆𝑐 leads to a more tractable form for the steady state operation.

Stephen Coward - April 2020 151

 Evaluation of metallurgical recovery factors for diamonds recovered from kimberlites

∫ 𝑵⋁𝝍(𝒙)𝒅𝒙 = ∫ 𝑵𝝍(𝒙)𝒖. 𝒏𝒅𝝈 Equation 19

𝑹𝒄 𝑺𝒄

which, when expanded, gives the working equation for the steady state operation
as:

𝑵 ∫ ⋁𝝍(𝒙)𝒅𝒙 + 𝑫 + 𝑩 = −𝑸 + 𝑨
𝑹𝒄

𝓡(𝝍(𝒙), 𝒙, 𝑭[𝝍(𝒙)])
𝑵 ∫ ⋁𝝍(𝒙)𝒅𝒙 + ∫ 𝒅𝒙
𝑹𝒄 𝑹𝒄 𝒎̅ (𝒙)
𝟏
+∫ ∫ 𝓡(𝝍(𝒙′ ), 𝒙′ , 𝑭[𝝍(𝒙)])𝐛(𝒙; 𝒙′ )𝒅𝒙′𝒅𝒙
𝑹𝒄 ̅ (𝒙) 𝑹′(𝒙)
𝒎
Equation 20
𝑵
−∫ ∫ 𝝍(𝒙′ ), 𝝁(𝒙′ ), 𝑭[𝝍(𝒙)])𝐚(𝒙; 𝒙′ )𝒅𝒙′𝒅𝒙
𝑹𝒄 ̅
𝒎 (𝒙) 𝑹′′(𝒙)
𝑵
= − ∑ 𝑾𝒐𝒖𝒕 𝒋 ∫ ( ) 𝝍𝒐𝒖𝒕 𝒋 (𝒙)𝒅𝒙
𝒋 𝑹𝒄 𝑴

𝑵
+ 𝑾𝒊𝒏 ∫ ( ) 𝝍𝒊𝒏 (𝒙)𝒅𝒙
𝑹𝒄 𝑴

The form of Equation 20 can be simplified as a ‘integro-differential equation’ for

analytical purposes.

𝓡(𝝍(𝒙), 𝒙, 𝑭[𝝍(𝒙)])
𝑵⋁𝝍(𝒙)𝒅𝒙 + ∫ 𝒅𝒙
𝑹𝒄 𝒎̅ (𝒙)
𝟏
+∫ ∫ 𝓡(𝝍(𝒙′ ), 𝒙′ , 𝑭[𝝍(𝒙)])𝐛(𝒙; 𝒙′ )𝒅𝒙′𝒅𝒙
𝑹𝒄 ̅ (𝒙) 𝑹′(𝒙)
𝒎
𝑵
−∫ ∫ 𝝍(𝒙′ ), 𝝁(𝒙′ ), 𝑭[𝝍(𝒙)])𝐚(𝒙; 𝒙′ )𝒅𝒙′𝒅𝒙 Equation 21
𝑹𝒄 ̅
𝒎(𝒙) 𝑹′′(𝒙)
𝑵
= − ∑ 𝑾𝒐𝒖𝒕 𝒋 ∫ ( ) 𝝍𝒐𝒖𝒕 𝒋 (𝒙)𝒅𝒙
𝒋 𝑹𝒄 𝑴

𝑵
+ 𝑾𝒊𝒏 ∫ ( ) 𝝍𝒊𝒏 (𝒙)𝒅𝒙
𝑹𝒄 𝑴

152 Stephen Coward - April 2020

 Chapter 6: Process simulation and rock characteristics

In the preceding section the fundamental population balance model has been
reduced to an appropriate discrete form with the region 𝑅𝑐 representing an
appropriate particle class, and the material entering and leaving each size class is
calibrated using data acquired from existing processes.

6.4 Unit process models

The unit processes that have a material impact on estimating the release, recovery
and loss of diamonds in the processes of comminution, separation and recovery
have been adapted for use in the integrated evaluation model.

There are three main processes in the diamond process flow sheet: size reduction
achieved in several steps, a density separation and final diamond recovery using a
combination of x-ray and magnetic recovery. Models used to simulate or emulate
these processes are described briefly below.

Crushing models

There are several types of crushers that are used in the diamond industry, including
jaw, gyratory, short-head and high-pressure roll crushers. In each part of the
flowsheet different modes of crushing are used. The key to understanding the
overall liberation is however the sequence of size distributions that are achieved.

Rocks do not generally fracture by breaking of atomic bonds, King (2001)

demonstrates that rocks that have measured hardness of ~50mpa would exhibit a
fracture strength of over 6000mpa if this were the case. It is suggested in the Griffith
theory of breakage (Griffith, 1921) that when a rock is exposed to rapid impact
stress it initially deforms elastically and absorbs the applied energy as elastic strain.
The strain grows until the stress intensity at the most sensitive flaw exceeds the
Griffith criterion for crack growth. The crack then propagates very rapidly and may
bifurcate several times being driven by the release of the stored strain energy, until
it meets an edge of the particle, and the original particle splits apart. The origin,
growth and the pattern/directions of crack propagation are a function of rock
characteristics (e.g., the contents and texture of the rock) and the nature of the
energy that is imparted (speed of loading, magnitude of stress, difference in primary
and secondary stresses) on the rock. King (2001) also suggests that the term
‘hardness’ be used to describe a rock’s resistance to indentation and that fracture
toughness be used to refer to a rocks resistance to the propagation of fracture for a
given set of stress or strain conditions.

Measuring and describing the relationship between energy input and the degree of
fracture produced is not trivial. Some of the challenges include:

Stephen Coward - April 2020 153

 Evaluation of metallurgical recovery factors for diamonds recovered from kimberlites

• Accurate measurement of the energy used to generate breakage;

• Quantitative description of the breakage produced;
• Description of the limits of the relationships measured.

Depending on how these problems are addressed will, by and large, dictate the
limits and usefulness of any subsequent process modelling framework.

The earliest references reviewed show that initial work was aimed at developing a
relationship between energy input and fracture achieved. The difficulty in
describing these relationships lies in the complex non-linear relationship between
energy input into the crushing device and in some way measuring and quantifying
‘breakage’ and ‘fracture’ as the reduction ratio.

As described by Holmes (1957) the earliest breakage models include those by

Rittinger (1867), Kick (1885) and Bond (1943) (see Equation 22 to Equation 25)
and Whitten (1972). The forms of these models are extensions of the earlier
approaches and their formulation is described below. Holmes (1957) carried out an
analysis of data from several deposits and suggested that the relationship needed to
be defined in a way that would include different behaviours and interactions that
were exhibited by different machines and different rock types.

Rittinger:

𝟏 𝟏
𝑾 = 𝑲𝟏 [𝟏 − ] . 𝟑 Equation 22
𝑹 𝒂

Kick:

𝒍𝒐𝒈𝑹
𝑾 = 𝑲𝟐 [ ] Equation 23
𝒍𝒐𝒈𝟐

Bond:

𝟏
𝟏 𝟐 𝟏
𝑾 = 𝑲𝟑 [𝟏 − ( ) ] . 𝟏 Equation 24
𝑹
𝒂𝟐

154 Stephen Coward - April 2020

 Chapter 6: Process simulation and rock characteristics

Holmes:

𝟏 𝒓 𝟏
𝑾 = 𝑲𝟒 [𝟏 − ( ) ] . 𝒓 Equation 25
𝑹 𝒂

Where:

𝑊 = 100𝑟 . 𝑊𝑖 is a constant for one material and one machine.

𝑅 is the particle size reduction ratio, usually expressed as the median size of the
feed to the crusher and divided by the median size of the product.

𝑟 is a variable parameter that is related to the material’s reluctance to grind as a

function of the material’s size. Holmes (1957) refers to this parameter as the “Kicks
law deviation exponent, since it is a measure of the departure of the crushing system
from ideal Kicks law behaviour”.

𝑎 is a product grind size usually represented by the single side dimension measure
of the square aperture that 80% of the product post crushing would pass.

Rittinger’s theory suggested that the energy necessary to reduce particle size is
proportional to the increase in specific surface area. It is focused on modelling the
rupture of chemical and physical bonds in the material.

In 1880 Kick suggested an energy fracture relationship. It was translated from the
original by Stadler to suggest “the energy required for producing analogous changes
in configuration in geometrically similar bodies of equal technological state varies
as the volumes of these bodies.” Kicks law centred on the energy required to deform
the particle to its elastic limit. In 1957 Holmes came to the forefront of the discussion
where up until then the ‘laws’ of Kick, Bond and Rittinger were in use.

Holmes’ model begins with the consideration of the failure of a cube of rock of
dimension D. Using Hooke’s law and three assumptions Holmes demonstrated that
by expansion of multiple failure events it can be shown that the energy required to
reduce a unit weight of cubes of side D can be formulated as in Equation 26:

𝟑𝒌 𝑫−𝒓 (𝑹𝒓 − 𝟏)
𝑬𝒘 = Equation 26
𝝆. (𝟐𝒓 − 𝟏)

Where:

Stephen Coward - April 2020 155

 Evaluation of metallurgical recovery factors for diamonds recovered from kimberlites

𝜌 is the material density.

𝐷 is the dimension of one side of the starting cube.

𝑟 is the variable parameter of the rock type considered referred to by Holmes as the
“Kicks law deviation exponent.”

𝑘 is another parameter of the model that is used to describe the failure boundary
conditions.

Holmes used an additional fourth assumption to derive a more generic form as

shown in Equation 27. The first assumption is that the model is independent of
shape changes. The second is that it is possible to represent the progeny distribution
through a single number (not withstanding prior comments about bimodality) in
this case the representation is given by the P80 of the distribution in the form stated
in Equation 27.

𝒀 = 𝟖𝟎(𝑿⁄𝒂)𝒎 Equation 27

Where:

𝑌 is the cumulative percentage passing at size X.

𝑋 is the size for which the percentage passing is to be calculated.

𝑎 is the size at which 80% passes.

𝑚 is a parameter which varies to some extent with a, normally it is taken as 1⁄√2.

The third assumption is that the effects of boundary conditions should be

considered. This includes the size and shape of the particles entering the system and
the machine that is used to perform size reduction.

The fourth assumption implies that there will have to be some consideration of the
proportion of energy that is used to propagate factures vs the energy that is input
into the device. This should take the form of an efficiency factor 𝜂 and will be
different for different sizes of material broken.

With these four assumptions it is possible to reduce Equation 27 to Equation 28:

156 Stephen Coward - April 2020

 Chapter 6: Process simulation and rock characteristics

𝑹𝒓 − 𝟏 𝟏
𝑾 = 𝑲( ). 𝒓 Equation 28
𝑹𝒓 𝒂

Where:

𝑊 is the work input into a comminution machine in KWh/tonne to a crushing

machine reducing the feed size 𝐹 to a product size 𝑎.

𝑅 = 𝐹/𝐴 is a parameter which varies to some extent with 𝑎 is the size at which 80%
passes.

240𝑘′ 𝑚(1.25)(𝑚−𝑟)⁄𝑚 (𝑅 𝑟 −1)

𝐾 = 𝜂. (𝑚−𝑟).(2𝑟 −1)
where k’ includes constants that enable the
𝜌
implementation to account for the units of measure of particle size and of power
input.

𝑟 is a variable parameter that is related to the material’s reluctance to grind as a

function of the material’s size

𝑚 is a parameter which varies to some extent with a, normally it is taken as 1⁄√2.

Holmes then suggests that the parameters in the model should be considered as
‘engineering measures’ rather than laws. He also suggests that the constants used in
the models are as much a function of the rock properties as the machine properties.
This substantiates the requirement for a framework to formalise the relationships
between primary, response and process properties of any comminution system. The
main shortcomings of direct explicit rock fracture models include the inability to
directly access, measure and hence model the internal so-called ‘co-ordinate shifts,’
and that these internal shifts are the result of an interdependent, potentially non-
linear function of the material properties, the machine properties and operating
characteristics of the comminution device. A further challenge to the explicit
modelling framework, as discussed by King (2001), is that the energy will propagate
into the progeny in several ways and differently into different factions of the
progeny as it evolves through the comminution device.

Repeating breakage events using a controlled apparatus that loads the particle in
some way to failure of single particles of a single mineral of a selected size can
however give some indication of the overall relationship between input energy,
probability of failure at differing energy levels and final progeny distribution (this
of course has to assume that the nature of flaws in the parent particles are randomly
distributed and is very similar between the particles tested, and that sufficient
particles of a given size can be tested to demonstrate the ‘average’ energy fracture

Stephen Coward - April 2020 157

 Evaluation of metallurgical recovery factors for diamonds recovered from kimberlites

response). Tavares and King (1998) conducted such tests on several Taconite
Quartz, Sphalerite and Galena particles and demonstrate that the cumulative
distribution of energies required to cause fracture can be represented using a log
normal distribution and have represented this relationship as per Equation 29.

𝒍𝒏(𝑬⁄𝑬𝟓𝟎 )
𝑷(𝑬; 𝒅𝒑 ) = 𝑮 ( ) Equation 29
𝝈𝑬

With G being the cumulative Gaussian distribution and 𝐸50 being the median
fracture energy which can be said to vary with size as per the relationship described
in Equation 30:

𝝋
𝒅𝟎
𝑬𝟓𝟎 = 𝑬⋈ (𝟏 + ) Equation 30
𝒅𝒑 − 𝒅𝒑𝒎𝒊𝒏

The values of 𝑑0 and 𝜑 are material specific and have been measured for several
common minerals. 𝑑𝑝𝑚𝑖𝑛 is the size of particles produced and below this size
particles absorb energy but do not fracture. 𝐸⋈ is the median fracture energy for
large particles. In this study ‘large’ was deemed to be above 1cm. Using the models
provided by King (2001). it is possible to plot the relationship between the energy
input and the probability of failure shown in Figure 51.

Relationship of Energy and Probability of Fracture for

Selected Minerals
99.99
99.95 Mineral
Series5
99.9
99.8 Galena
Cumulative Probability of Fracture %

99.5
99 Sphalerite
98 Quartz
95 Taconite
85
80
70
60
50
40
30
20
15
10
5
2

0.5
0.2
0.1
0.05

1 10 100 1000
Log Specific particle fracture Energy (J/kg)

Figure 51: A plot of the specific input energy and cumulative probability of failure for a few
selected minerals, size parameter set to 5mm (Adapted after King 2001).

The impact of the size selected also exhibits an exponential form, this is shown in
the plot in Figure 52. This clearly demonstrates that the measurement and
estimation of the energy required to fracture smaller particles is not trivial and

158 Stephen Coward - April 2020

 Chapter 6: Process simulation and rock characteristics

especially as the material is ground ever finer the distribution of particle fracture
toughness will also expand dramatically.

Relationship of Size and E50 of Fracture for Selected

Minerals
800
Galena
700 Sphalerite
Median Fracture energy (J/Kg)

Quartz
600
Taconite
500

400

300

200

100

0
0 2 4 6 8 10 12
Parent Particle Size in mm

Figure 52: A plot of the median fracture energy for mineral particles of different sizes.

Given the distributions shown in Figure 51 and Figure 52 it is evident that even for
a mineral with a defined composition the range of energy required to achieve the
same degree of breakage varies substantially. This suggests that it would be limiting
to consider a single value for rock toughness per lithology when modelling the
operation of process plant equipment. Rather it is proposed that a variable that
describes the range of 'resistance to breakage' or so-called 'toughness', should be
spatially estimated into the block model. This would describe the rock properties in
a block to be mined and become a variable that can be used as an input into the
process model. This requires a review of approaches to population balance
modelling to determine how the variability and uncertainty in rock toughness can
be realistically incorporated into these models.

The general equation, according to King (2001), for the population balance model
can be written for comminution machines as follows:

Stephen Coward - April 2020 159

 Evaluation of metallurgical recovery factors for diamonds recovered from kimberlites

𝒅 𝓡(𝝍(𝒙), 𝒙, 𝑭[𝝍(𝒙)])
𝑵 (𝒖(𝒙)𝝍(𝒙)) +
𝒅𝒙 𝜷𝒙𝟑
𝟏
+ 𝟑 ∫ 𝓡(𝝍(𝒙′ ), 𝒙′ , 𝑭[𝝍(𝒙)])𝐛(𝒙; 𝒙′ )𝒅𝒙′𝒅𝒙
𝜷𝒙 𝑹′(𝒙)
𝑵 ′ ), ′
𝒅𝜷𝒙′𝟑
− 𝟑∫ 𝝍(𝒙 𝝁(𝒙 ), 𝐚(𝒙; 𝒙′ )𝒅𝒙′𝒅𝒙
𝜷𝒙 𝑹′′(𝒙) 𝒅𝒙 Equation 31

𝑵
= − ∑ 𝑾𝒐𝒖𝒕 𝒋 ∫ ( ) 𝝍𝒐𝒖𝒕 𝒋 (𝒙)𝒅𝒙
𝒋 𝑹𝒄 𝑴

𝑵
+ 𝑾𝒊𝒏 ∫ ( ) 𝝍𝒊𝒏 (𝒙)𝒅𝒙
𝑹𝒄 𝑴

In Equation 31 the scalar 𝑥 represents the particle size 𝑑𝑝 measured in linear mm

and the average mass of the particle can be related to size by Equation 32:

̅ (𝒙) = 𝜷𝒙𝟑
𝒎 Equation 32

A commonly used mode for abrasion rates is to assume that it is proportional to the
surface area. It is then possible to relate the change in the proportion of material in
size 𝑑𝑝 per unit time as some function of the area of the particle (Equation 33).

𝝅 𝒅𝒙𝟑 𝒌′𝝅𝒙𝟐
𝝅 =− Equation 33
𝟔 𝒅𝒕 𝟐

This suggests that the rate at which particles move in phase space is considered
constant and can be given by (Equation 34):

𝒅𝒙
𝒖(𝒙) = = −𝒌′ Equation 34
𝒅𝒕

And hence more generally as in Equation 35:

𝒅𝒙
= −𝒌(𝒙) = −𝒌𝒙𝚫 Equation 35
𝒅𝒕

If delta is a constant varying between 0 and 1 then the rate of change of mass with
respect to time is given by:

160 Stephen Coward - April 2020

 Chapter 6: Process simulation and rock characteristics

𝒅𝒎 𝝅𝝆𝒔 𝒙𝟐+∆
= −𝒌 Equation 36
𝒅𝒕 𝟐

It is then possible to convert Equation 31 above using the relationship to derive

Equation 37:

𝑴 𝒑(𝒙)
𝝍(𝒙) = Equation 37
𝑵 𝜷𝒙𝟑

And so, the population balance equation can be converted to (Equation 38):

𝑴 𝒅(𝒌(𝒙)𝒑(𝒙)/𝒙𝟑 𝓡(𝒑(𝒙), 𝒙, 𝑭[𝒑(𝒙)])

− +
𝒃 𝒅𝒙 𝜷𝒙𝟑
𝟏
+ 𝟑 ∫ 𝓡(𝒑(𝒙′ ), 𝒙′ , 𝑭[𝒑(𝒙′)])𝐛(𝒙; 𝒙′ )𝒅𝒙′𝒅𝒙
𝜷𝒙 𝑹′(𝒙)
Equation 38
𝑴 𝒑(𝒙′ )
− 𝟑∫ 𝟑𝒌(𝒙′ )𝒙′𝟐 , 𝐚(𝒙; 𝒙′ )𝒅𝒙′
𝜷𝒙 𝑹′′(𝒙) 𝒙′𝟑
𝒑𝒐𝒖𝒕 (𝒙) 𝒑𝒊𝒏 (𝒙)
= −𝑾𝒐𝒖𝒕 𝒋 𝟑
+ 𝑾𝒊𝒏
𝜷𝒙 𝜷𝒙𝟑

There are several unit process models in use in the industry that are based on the
principles described above; these include models by Whitten (1972) and
Andersen(1988).

The Whitten breakage and selection functions (Napier-Munn, 1999) can be

described as a combination of models for selection and fracture. In each size class
the number of particles leaving and entering the size class is defined for each
breakage ‘event.’ The selection and breakage functions can be calibrated for each
ore type. Equation 39 presents a formulation of this approach:

𝒏 𝒏

𝑷𝒊 = 𝒇𝒊 + 𝝉 ∑ ∑ 𝒃𝒊𝒋 . 𝑪𝒋 . 𝑷𝒋 − 𝝉 𝑺𝒊 𝑷𝒊 Equation 39
𝒊=𝟏 𝒋=𝟏

Where 𝑃𝑖 is the mass proportion of material in class i

𝑓𝑖 is the feed to class 𝑖

Stephen Coward - April 2020 161

 Evaluation of metallurgical recovery factors for diamonds recovered from kimberlites

𝑏𝑖𝑗 is the fraction of material reporting to size class j when material in size class i is
comminuted

𝜏 is the time step

𝐶𝑗 is the composition class j

𝑃𝑗 is the proportion of particles of size i in composition class j

𝑆𝑖 is the selection rate for breakage of particles in class i

This can be rearranged as follows in Equation 40:

𝒏 𝒏

𝑷𝒊 (𝟏 − 𝝉𝑺𝒊 ) = 𝒇𝒊 + 𝝉 ∑ ∑ 𝒃𝒊𝒋 . 𝑪𝒋 . 𝑷𝒋 − 𝝉 𝑺𝒊 Equation 40

𝒊=𝟏 𝒋=𝟏

And thus, the proportion within each class can be determined using Equation 41:

𝒇𝒊 + 𝝉 ∑𝒏𝒊=𝟏 ∑𝒏𝒋=𝟏 𝒃𝒊𝒋 . 𝑪𝒋 . 𝑷𝒋 − 𝝉 𝑺𝒊

𝑷𝒊 = Equation 41
(𝟏 − 𝝉𝑺𝒊 )

The formulae are, however, not in a closed form and must be solved iteratively
across all sizes and for each class within each size fraction. The factors that are used
to define the parameters for breakage rate and selection function are often derived
through experimentation. Often a statistic, commonly the means, of values derived
from an assessment of comminution performance of several batch tests on small
samples are used in the calibration. This approach requires an assumption that the
samples are representative of the material that is to be treated in production, and
that the change in scale between laboratory scale machines to full scale operational
machines will not result in a substantial bias.

To use this function in the ‘Integrated Evaluation Model’ the breakage and selection
functions need to be fitted to each comminution device in the process flowsheet and
must also be calibrated for material domains in the kimberlite pipe.

An Application of the Whitten Crusher model

A feed with a distribution as described in column fi of Table 23 is sent to a

comminution device. The selection function is described by column Si and the time

162 Stephen Coward - April 2020

 Chapter 6: Process simulation and rock characteristics

adjustment factor (T) is derived for each size fraction, to determine the percentage
of each size fraction that will be selected for breakage. The blue block in this table is
the breakage matrix or breakage function. It describes the post-crushing
distribution of material selected and crushed from each size fraction. The
distributions in the breakage matrix can be multiplied by the proportional fraction
of each size that is selected for breakage to determine the combined size distribution
of material that is selected and broken. This distribution can be added to the
material that was not selected for breakage and passed through the unit unscathed.
In this way the crushed distribution can be determined, shown in the last row of the
table.

Mins Frac of Hour

Residence time 30 0.5
Destination Sizes - breakage matrix
% per Actually
Class Size class in Selection Time selected 100 10 3.16 1.78 1.33 1.15 1.07
midpoint Class the feed function Adustment per class
mm Num fi Si T % mm mm mm mm mm mm mm
100.00 1 10 0.5 0.5 2.00 0.1 0.1 0.1 0.1 0.3 0.3
10.00 2 15 0.3 0.5 1.96 0.1 0.1 0.2 0.3 0.3
3.16 3 25 0.2 0.5 2.27 0.1 0.2 0.3 0.4
1.78 4 12 0.4 0.5 2.00 0.1 0.5 0.4
1.33 5 16 0.7 0.5 4.15 0.5 0.5
1.15 6 12 1 0.5 4.00 1
1.07 7 10 0 0.5 0.00 1
Total 100 Total 16.38 0.1 0.2 0.3 0.6 1.9 3.9
Total
Distribution of material selected and crushed 0.0 0.2 0.4 0.6 1.2 4.9 9.0 16.4
Unselected in each class 8.0 13.0 22.7 10.0 11.9 8.0 10.0 83.6
New Distribution 8.0 13.2 23.1 10.6 13.1 12.9 19.0 100.0

Table 23: A tabular demonstration of a Whitten selection and breakage model.

The calibration of this model requires several inputs to determine the functions that
underpin selection and breakage. As the material or the unit operating parameters
change, these functions will require recalibration. The model does not have a direct
link to material properties but does provide a large degree of flexibility to model a
range of comminution devices. The derivation of the different breakage functions
for each size has been the focus of a large body of work. More recent work carried
out at the JKMRC has described a form of energy breakage modelling that is based
on measurements made on single particle fracture tests (Napier-Munn et al., 1999).
The framework for this is expanded below, as it provides a way to relate material
properties to expected size distribution of the rock and hence to the diamonds that
will be liberated.

Energy Fracture modelling

One useful form of the comminution model is based on the idea that the progeny of
breakage events can be based on a mixture of two separate populations and that
each cumulative population can be modelled in the form depicted in Equation 42.

Stephen Coward - April 2020 163

 Evaluation of metallurgical recovery factors for diamonds recovered from kimberlites

𝒙 𝒏
𝑩(𝒙; 𝒚) ∝ ( ) Equation 42
𝒚

where:

𝐵(𝑥; 𝑦) is the breakage function

𝑥
is the ratio of the size of the progeny to the size of the original particle
𝑦

𝑛 is used to describe different distributions of particles produced from tensile stress

and particles produced by compressive stress at the points of contact in the
comminution device. These two distributions are added using a suitable weighting
scheme.

𝒙 𝒏 𝒙 𝒏𝟐
𝑩(𝒙; 𝒚) = 𝑲 ( ) + (𝟏 − 𝑲) ( ) Equation 43
𝒚 𝒚

The first term in Equation 43 describes the distribution of particles generated by

compressive stress at the point of contact, and the second term describes the
progeny of tensile stress fracture. It is possible to determine the constants K, 𝑛 and
𝑛2 experimentally for a given rock type and fracture mechanism.

The data gathered from drop weight tests in which a single particle is subject to a
known input energy can be used to determine the size distribution of fine and coarse
particles. These can be plotted on log-log plot of the size of progeny vs the breakage
function value. King (2001) suggests that if 𝑑𝑟 is a representative size for a
distribution, say the size for which 80 per cent of the particles are smaller than, then
a change in the breakage energy depends inversely on the initial particle size.

It is possible to describe a generic form of relationships demonstrated by Kick

(1883), Rittinger (1857) and Bond (1943) as shown in Equation 44 :

𝒇(𝒅𝒓 ) = −𝑲𝒅−𝒏
𝒓 Equation 44

Where

𝐾 is a constant

164 Stephen Coward - April 2020

 Chapter 6: Process simulation and rock characteristics

𝑛 is a constant for which a value above 1 reflects an increase in energy per

unit mass as the particle size decreases.

It is possible to integrate Equation 44 to determine the energy that will be required

to change the representative size. Starting with the assumption that 𝐸 = 0 for the
initial distribution described by 𝑑𝑟𝑖 , this derivation is shown in Equation 45

𝒇(𝒅𝒓 ) = −𝑲𝒅−𝒏
𝒓

𝑲
𝑬 = − 𝟏−𝒏 𝒅𝟏−𝒏
𝒓 +𝑪 𝒏≠𝟏

Equation 45

𝑲
𝟎=− 𝒅𝟏−𝒏 + 𝑪
𝟏−𝒏 𝒓

𝑲 𝟏 𝟏
𝑬= ( (𝒏−𝟏) − (𝒏−𝟏) )
𝟏−𝒏 𝒅 𝒅
𝒓 𝒓

This form can be used to show how the equation changes for values used by Kick
where n=1, Bond where n = 1.5, and Rittinger where n=2 (Daniel, Lane and McLean,
2010). Single impact tests can be used to determine the progeny distribution that is
achieved for a given energy input. The data produced can then be plotted as the
cumulative percentage passing a given size on a relative size scale.

The distribution of the progeny can be characterised by the percentage passing one-
tenth of its original size (𝑡10 ). This is another form of the representative size 𝑑𝑟
referred to above. In the prior work this size was defined as the 𝑃80 . It is suggested
that given 𝑡10 it is possible to define a relationship that will describe the other points
on the distribution (𝑡𝑛 ) . The value for the 𝑡10 is determined experimentally and is
derived from Equation 46:

Stephen Coward - April 2020 165

 Evaluation of metallurgical recovery factors for diamonds recovered from kimberlites

𝒕𝟏𝟎 = 𝒕𝟏𝟎 𝒎𝒂𝒙 (𝟏 − 𝒆−𝜷𝑬⁄𝑬𝟓𝟎 ) Equation 46

Where

𝑡10 𝑚𝑎𝑥 is a material specific breakage function;

𝛽 is a material specific breakage parameter;

𝐸50 is the median fracture energy.

Once the value of 𝑡10 has been derived then the rest of the expected progeny size
distribution for a specific energy input to a specific material can be determined. This
is often done with truncated distribution functions such as the Rosin-Rammler
described in Equation 47.

𝟏𝟎−𝟏 𝜶
( )
𝒕𝒏 = 𝟏 − (𝟏 − 𝒕𝟏𝟎 ) 𝒏−𝟏 Equation 47

Where

𝑡𝑛 is the cumulative percentage of progeny passing a proportional size;

𝑡10 is material specific breakage parameter;

𝛼 is material specific constant; and

𝑛 is the fractional size number.

It is possible to plot the expected size distribution for a range of progeny sizes using
the above relationships. A worked example from King (2001) is depicted in Figure
53.

166 Stephen Coward - April 2020

 Chapter 6: Process simulation and rock characteristics

10.00

Cumulative % passing =Brekage Function

1.00

apatite 6488

0.10 apatite 1504

apatite 874
apatite 245
0.01
0.01 0.1 1

0.00
Relative Particle size d/dp'

Figure 53: A plot of energy input vs product size using the t10 approach and a Rosin Rammler
breakage function modified after King (2001).

Screen models

These models are used to determine the split of material through a screening device
and have an impact on the ultimate size distribution of material that either leaves
the circuit or that is fed to the next crushing or comminution unit process.

There are several versions of the model, but all are based on determining the
probability of particles of a given size being presented to an aperture and passing
the given screen aperture. The model calibration varies depending on the shape of
the aperture, the loading of the screen and the screen deck inclination.

𝑷 = 𝑲[(𝒏 − 𝟏)⁄𝒏]𝟐 Equation 48

In Equation 49 the probability 𝑃 of a particle of dimension 𝑙⁄𝑛 passing through a

square aperture of dimension 𝑙 is described by the squared ratio of (𝑛 − 1)⁄𝑛
where 𝑛 is any number greater than 1. K is a process-specific parameter used to
adjust for operational inefficiency and calibrated by direct experimentation
(Taggart, 1964).

Stephen Coward - April 2020 167

 Evaluation of metallurgical recovery factors for diamonds recovered from kimberlites

DMS Models

Estimating separation efficiency and forecasting the yield tonnage is a very

important component of the flowsheet simulation. The DMS process is responsible
for the highest proportion of material mass reduction and hence errors in this
process model will have a marked impact on the overall recovery model accuracy.
In its simplest form the yield can be predicted by taking a sample from the orebody
and determining the proportion of material that will be in each density class and
multiplying this by the probability of material in a density class reporting to the sink
or recovered fraction stream. This curve is commonly referred to as a partition
curve. The parameters of this curve are usually determined experimentally, and
these data are used to derive the Ep value as described in Equation 49:

𝑫𝟕𝟓 − 𝑫𝟐𝟓
𝑬𝒑 = Equation 49
𝑫𝟓𝟎

Where:

Ep is the probability of separation.

D75 is the density at which 75% of the material reports to the sink fraction.

D50 is the density at which 50% of the material reports to the sink fraction.

D25 are the densities at which 25% of the material reports to the sink fraction.

There is also an option to include a constant K value in the estimation of separation

efficiency that reflects the proportion of material that is misplaced due to its shape
or other inefficiencies. To determine the proportion of material that will report from
any given density fraction to the sinks, the formula above is rearranged to give the
following Equation 50:

𝑬𝒑
𝑷𝒔 = 𝑲×(𝑫 Equation 50
𝟏− 𝒆 𝒄 −𝑫𝟓𝟎 )

Where:

168 Stephen Coward - April 2020

 Chapter 6: Process simulation and rock characteristics

Ps is the probability of sink.

Ep is the probability of separation.

Dc is the density of the class being considered.

K is a separation efficiency factor.

Estimating the proportion per density class per size fraction for a given kimberlite
size distribution is not trivial. The laboratory process for densiometric
determination typically consumes about 70kg of core that is crushed in a jaw
crusher to produce a given size distribution, and then the product is screened into
several size fractions. Each size fraction is then split into density classes using
liquids of ascending density. The products from each sink float split are weighed and
documented. This produces data that can be represented in a table such as in Figure
54.

Proportion of Material By Density Class

0.5
Proportion By Weight in Desnity Class

0.45
0.4
0.35
0.3
(Percent)

0.25
0.2
0.15
0.1
0.05
0
Sample 1 Sample 2 Sample 3 Sample 4

2.5 2.9 3 3.1 3.2

Figure 54: A Densimetric distribution plotted for four samples derived from kimberlite,
crushed to 100% passing 12mm and grouped by density classes.

• It is, however, evident that as the size distribution changes the distribution
of density within each size class will change. The adaption of the density
distribution for changing size distribution requires a function that relates the
change in size distribution to the change in density distributions in each size
class. If one were to consider five density classes and 13 size classes one
would end up with 75 'density by size' classes, so material breaking out of
one of the larger size classes will produce progeny that has some undefined

Stephen Coward - April 2020 169

 Evaluation of metallurgical recovery factors for diamonds recovered from kimberlites

probability of ending up in several of the smaller density classes. To do this

mapping requires staged comminution and density test work for each
material type considered. The number of tests and stages of crushing
required to facilitate this multidimensional mapping depend to a large
degree on:
• the density distribution of the components that make up the rock;
• size distribution of different components; and
• the affinity or adherence between the different components in the rock
under investigation that will give rise to the composites at different levels of
fracture

This approach can to some extent be informed by the texture of the kimberlite rock
type and will also provide a quantitative approach to classification of textures that
result in distinctly different response variables.

6.5 Establishing relationships between comminution and diamond

recovery

Once it is possible to relate rock property to process performance, and especially

total comminution that will be achieved at a block scale, a model of diamond
liberation and lock up is required to enable the estimation of a diamond recovery
factor. The next section describes in some detail how the particulate nature, and size
distribution relationships that have been shown to exist in kimberlitic diamond
deposits can be used to relate overall comminution and DMS efficiency to diamond
recovery.

The macro micro diamond recovery model

In the earliest phases of a diamond project the grade is estimated using

microdiamonds, recovered from small samples(~100-400kg) that generally include
diamonds smaller than ~ 0.7mm. These are recovered from core in a process of
thermochemical dissolution and evaluated by taking microscopic measurements of
each diamond. The estimate is achieved by determining the abundance of small
diamonds in several size classes, and then modelling the diamond size frequency in
the form of a quadratic equation as depicted in Equation 51.

𝒚𝒊 = 𝒂𝒙𝟐 + 𝒃𝒙 + 𝒄 Equation 51

This curve model can then be extrapolated into the larger diamond size classes. This
is commonly referred to as a total content curve (see Figure 55 below).

170 Stephen Coward - April 2020

 Chapter 6: Process simulation and rock characteristics

Grade-Size Distribution Stones per 100tons per unit size interval -

Sample data and model

1,000

Number of Stones
10

0
0.00001 0.0001 0.001 0.01 0.1 1 10 100

Weight in carats

Sam ple Grade spht/ui Estim ated spht/ui

Figure 55: A plot of the log of diamond weight vs log of the number of stones in each class
per hundred tonnes per unit interval.

The total stone content can be derived by integrating this function across all size
fractions (Equation 52). This can be used to determine the grade between an upper
and lower cut off size by converting the stones frequency to a mass.

𝒅𝒏
𝒕𝒐𝒕𝒂𝒍 𝒄𝒐𝒏𝒕𝒂𝒊𝒏𝒆𝒅 𝒔𝒕𝒐𝒏𝒆𝒔 = ∫ (𝒚)𝒅𝒙 Equation 52
𝒅𝒊

Process plants however usually operate in the size range of 1mm to 32mm although
some plants operate beyond this size range (Technical and Financial Report, 2001).
To adapt this curve to estimate recoverable grade there is a requirement to trim the
upper and lower ends of the curve. This is done by fitting a third-order polynomial
to the total content curve at both the top and bottom end of the planned recovery
size envelope. The shape and sharpness of the top and bottom size cut-off curves is
usually determined by the modeller and is based on experience of curves achieved
at several existing operations.

Stephen Coward - April 2020 171

 Evaluation of metallurgical recovery factors for diamonds recovered from kimberlites

Total diamond content

5 tonne sample, +1mm recovery

16
5 tonne sam ple
14

12 Model 5T sam ple

10

8
Grade Ln(Spht/ui)

-2

-4

-6

-8
-8 -6 -4 -2 0 2 4 6 8
Size Ln(Carats)

Figure 56: A plot of the logarithm of diamond size vs stone frequency in stones per hundred
tonnes per unit interval.

This trimming can then be used to determine the resource grade and size
distribution that is reported at a bottom size cut-off of 1mm. In cases where
economics suggest that the lower cut-off should be higher than 1mm the curve is
adjusted to the recommended cut-off in a similar way to that shown in Figure 56

Benefits and limitations of the micro - macro modelling approach

The sampling for microdiamonds requires of the order of 50 to 400kg of material

that is digested in a sequence of chemical treatments that do not require any
crushing and do not leave any residue larger than several microns. This process
does not damage the contained diamonds and does not leave any diamonds locked
in the kimberlite. With sufficient samples it is possible to distinguish size
distributions within internal lithology types in kimberlite deposits. The method,
however, relies on extrapolating from the microdiamond size range to the macro-
diamond size range. This extrapolation becomes subjective in the absence of an
abundance of macro-diamonds that can be obtained from bulk sample treatment.

Samples that provide macro-diamonds need to be in the order of tens to hundreds

of tonnes and may, or may not, be representative of the diamond population in each
of the lithologies in the deposit. Also, the recoveries from bulk sampling are subject
to loss and lock up that is a function of the total grind achieved and the efficiency of
the diamond processing equipment.

172 Stephen Coward - April 2020

 Chapter 6: Process simulation and rock characteristics

The trimming of the ends of the curves is also subjective in this method in the
absence of operating plant data, and small increments in the larger sizes have a
relatively low impact on the carat recovery but can have a substantial impact on the
revenue recovery. The benefits of having an underlying diamond size frequency
model when sampling for grade are described in Coward and Ferreira (2004)

The granulometry liberation and lock-up model

Once the in situ total content has been estimated and relatively strict cut-off sizes
applied, the process of diamond recovery and loss can be modelled in more detail.
As diamonds are particulate in nature, having a large range of size distributions, one
of the ways in which the estimated diamond recovery can be modelled is to relate
the comminution of rock to the expected proportion of diamonds that will be
released or liberated from the kimberlite.

The model was originally conceived and applied to operating facilities on the west
coast of Namibia with the aim of trying to calculate the loss of diamonds that arises
from the inability to crush all the material to lower than the smallest contained
diamond (Kleingeld, 1982). Diamonds lost in this way are deemed to be "locked".
The model is centred on an assumption that it is possible to build a relationship
between the recovered size distribution of diamonds and the grind size of the ore
achieved to estimate the diamonds that are still not liberated, and hence discarded
in the tailings. The model was subsequently applied by Ferreira to a number of
kimberlitic deposits. During this time a collaboration with Lantuéjoul(1998) saw
several improvements in the formulation of the model. This author worked with
Ferreira from 1998 to 2008 to improve various aspects of sampling, data collection
and its use in implementing the model at several of De Beer's operating mines.
Ferreira (2013) gives a brief description of the model, however a more detailed
description is presented here due to its importance in this research. The model can
then be used in a variety of ways with total content curves to estimate the expected
recovered diamond distribution.

Model description

The model begins with an assumption that the diamonds that have been recovered
represent only a portion of the total population of diamonds. The total population
of diamonds is deemed to be the sum of the diamonds recovered and the diamonds
that are still ‘locked’ in the discarded tailings. The model assumes that there is a
relationship between the proportion of recovered diamonds within a size class and
the proportion of the discarder kimberlite in an equivalent size class. An additional
assumption of the model is that the that the diamonds are not damaged in the rock
crushing process.

Stephen Coward - April 2020 173

 Evaluation of metallurgical recovery factors for diamonds recovered from kimberlites

A very simple conception of the model can be applied to a process that only includes
crushing of the rock and handpicking out the diamonds. Assume that prior to
crushing all the rock particles were in one size class, say 5mm to 4mm and that the
deposit only contains diamonds in the 5mm to 4mm fraction. If we do not crush the
ore any further and carry out a hand sorting , i.e. there are no ore particles less than
the 5-4mm size class, it can be reasonably assumed there is 0% liberation. If,
however the grind achieved meant that there was still 50% of the material in this
fraction, and the other 50% was less than 4mm it would be possible that only half of
the diamonds were liberated. By extension if the grind achieved ensured that there
was no material larger than 4mm, and none of the diamonds were damaged in the
kimberlite size reduction process, then all the diamonds would be liberated.

The range of sizes of diamonds recovered and the range of size of ore particles
(grind) produced can be obtained by sampling and screening the discarded ore, and
screening and recording the mass size distribution of the diamonds recovered. The
relationship between the size distribution of the recovered and the kimberlite
discarded can be used to develop an estimate of the diamonds that are still
contained in the discarded kimberlite stream.

There are several sampling and data assumptions required by the model:

• It is possible to define the relationship between diamond size classes and ore
size classes;
• It is possible sample for, and reconstitute the total grind of the diamond
bearing ore;
• The diamond concentration is relatively low and hence the crushed size
distribution of the rock is independent of the size distribution of the
diamonds;
• Diamonds are randomly distributed in the ore blocks; and
• There is no relationship between the diamonds size and its location in
broken particles.

Several ore sizes are tabulated in Table 24. From the recovery process the total grind
of the ore fed to the plant is recorded. This size distribution is reflected in the third
column of this table. The recovered diamonds could be sieved, and the size
distribution represented in the fourth column as the percentage of stones per class.
By summing the classes from the smallest to the largest it is possible to express the
proportion of the distribution that could still be “locked” within each ore size class.

174 Stephen Coward - April 2020

 Chapter 6: Process simulation and rock characteristics

Measured Data Model Calculation

% of Diamond
Class Cumulative % of
Class % of Ore in Size % of Diamonds in Distribution Cts Locked per
bottom Diamonds Less
Number Class Size Class Possibly Locked in 100cts Recovered
Size than Size Class
Class
1 15 10 0 100 100 10
2 10 20 15 85 85 17
3 8 50 35 50 50 25
4 4 15 45 5 5 0.75
5 2 5 5 0 0 0
Totals 100 100 52.75

Table 24: Calculation of locked diamond content.

It is then possible to calculate the proportion of carats that can be locked in each size
fraction by multiplying the proportion of the ore that is in each size class and the
proportion of the diamond size distribution that could be contained in the ore
particle. As depicted in Table 24 the second class (10mm) of ore particles make up
20% of the material discarded. 15% of the total diamonds recovered lie in this class,
and 85% of the diamonds recovered are below 10mm. This suggests that 85% of the
diamond population can be locked, and by multiplying the ore proportion by the
proportion of the diamond distribution that can be locked we calculate that the lock
up in this size class can be 17 cts. By summing the locked content estimated for each
size fraction we see that for every 100cts recovered 52.75 cts could be locked. This
implies that the liberation efficiency can be estimated to be:

𝟏𝟎𝟎
%𝑳𝒊𝒃𝒆𝒓𝒂𝒕𝒊𝒐𝒏 = 𝑿𝟏𝟎𝟎 Equation 53
𝟏𝟎𝟎 + 𝒍𝒐𝒄𝒌𝒆𝒅 𝒆𝒔𝒕𝒊𝒎𝒂𝒕𝒆𝒅

So, in this case the liberation would be expected to be 65%.

The process, however, also includes a float and sink separation process, and for a
kimberlite particle to sink and be recovered in the final process, the density of the
particle must exceed that of the effective cut point in the dense media separation
process. This means that is possible to calculate the maximum size of diamond that
can be floated out of the dense media separation process given the ore density,
diamond density and the density cut point in the plant. This change in the lock up
calculation by limiting the maximum size of diamond that can be recovered to below
the ore size is depicted in table 25.

Stephen Coward - April 2020 175

 Evaluation of metallurgical recovery factors for diamonds recovered from kimberlites

Measured Data Model Calculation

% of Diamond
Class Cumulative % of Maximum Size of Cts Locked
Class % of Ore in Size % of Diamonds in Distribution
bottom Diamonds Less Diamond in this per 100cts
Number Class Size Class Possibly Locked in
Size than Size Class Class Recovered
Class
1 15 10 0 100 10 85 8.5
2 10 20 15 85 8 50 10
3 8 50 35 50 4 5 2.5
4 4 15 45 5 2 0 0
5 2 5 5 105 1 0 0
Totals 100 100 21

Table 25: Calculation of lock up with constraint placed on the maximum size of contained
diamond.

In this adapted calculation, the class 2 (10mm) ore particles can only lock up
diamonds up to 8mm. There are only 50 % of diamonds below this size limit that
can potentially be locked. The same calculations are carried out. The estimated
locked content has dropped to 21cts for every 100cts recovered, equating to a
liberation of 83%.

To generate a model requires a realistic method for determining the limiting mass
of diamond that can be contained in a particle and for that particle to be recovered
by the process. One approach to derive this limit is to consider the physical nature
of the process used to separate diamonds from kimberlite post the crushing stage.
In this process the density differences between diamonds (specific gravity of 3.52
g/cm3) and most kimberlites (specific gravity of 2.7 g/cm3) is exploited to generate
a diamond rich concentrate. The maximum volume of diamond contained in a
particle of kimberlite that could be floated out of the dense media separation
process is a function of the density of media that is used to separate the diamonds
from the kimberlite, the density of the kimberlite, the density of the diamond and
the relative volume of the diamond and kimberlite in the particle. A critical particle
can be defined as one where the forces of buoyancy and the sinking force are equal.
The mass of displaced medium will equal the mass of the particle, and the mass of
the particle is the combined mass of the contained diamond and the enclosing
kimberlite.

Model assumptions

The main underlying assumptions of the liberation model include:

• Total diamond content is the sum of the recovered and locked

diamonds;
• The density of kimberlite particles is homogenous across a range of
sizes and particle shapes;
• The diamonds exist in very low concentrations and thus have little
influence on the location and orientation of fractures in the rock;

176 Stephen Coward - April 2020

 Chapter 6: Process simulation and rock characteristics

• All diamonds that report to the sink stream, including those in

composite diamond and kimberlite particles, will be recovered;
• The recovered diamond size distribution and the in situ size
distribution do not differ markedly; and
• The diamond size distribution can be reliably modelled as a log
normal distribution.

Model formulation

The general formulation of the liberation model will be described in this section and
is broadly based on the work of Kleingeld (1982) and Ferreira (2013) and
unpublished work by Lantuéjoul(1998).

The notation used is described in King (2000). The particle size distribution function
𝑃(𝑑𝑝 ) is the mass fraction of that portion of the size fraction that consists of particles
of a size less than or equal to dp where dp is the size of the particle. This function has
the properties:

P (0) =0
P (∞) =1
P(dp) increases monotonically from 0 to 1 as dp increases from 0 to ∞

By differentiating the distribution function, it is possible to work out proportions of

particles that are less than or equal to a given size threshold x. The distribution
density function p(x) can be defined by Equation 54:

𝒅𝑷(𝒙)
𝒑(𝒙) = Equation 54
𝒅𝒙

And likewise, the discrete density distribution can be related to its density function
in a similar way (Equation 55):

𝑫𝟏−𝟏

𝒑𝒊 = ∑ 𝒑( 𝒙)𝒅𝒙 Equation 55
𝑫𝒊

The usefulness of this approach for describing particle size distributions is that it
facilitates the assumption that it is possible to work with the size class rather than
the individual particles. To do this requires the definition of the ‘average’ particle in
a size class, and the allocation of several empirical distribution functions to the
particle populations.

Stephen Coward - April 2020 177

 Evaluation of metallurgical recovery factors for diamonds recovered from kimberlites

In reality the distribution for a sample of ore or diamond particles is truncated at

both the upper end, due to the maximum size of rock that is considered, and at the
lower end by the minimum size of particle that is sampled or diamond that is
recovered. The continuous distribution could be obtained for a parcel of diamonds
by weighing each stone. It is, however, normal practice for the particles to be sieved
on a set of screens and for the mass of particles in each size class to be measured.
This then gives rise to the discrete particle size density function (Equation 56):

𝐃𝐢−𝟏
𝐏𝐢 (𝐝𝐩 ) = ∫ 𝐝𝐏(𝐝𝐩 ) = 𝐏(𝐃𝐢−𝟏 ) 𝐏(𝐃𝐢 ) = ∆𝐏𝐢 Equation 56
𝐃𝐢

Pi is the mass fraction of the particle population that consists of particles between
size Di and Di-1.

∆𝑑𝑝 = 𝐷𝑖−1 − 𝐷𝑖 is known as the size class width, with the upper and lower-class
size boundaries given by Di and Di-1 respectively. The representative size in this class
is required so that the ‘average’ characteristic in this class can be used in modelling
the behaviour of all particles within this class. One method proposed by King (2001)
is to develop the number density distribution function (as opposed to a mass
distribution considered above). The number distribution function for any
characteristic can be defined as𝜑(𝑥), which is the faction by number of particles in
the population having size equal to 𝑥 or less. The number density function can be
defined as in Equation 57:

𝒅𝝋(𝒙)
𝝋= Equation 57
𝒅𝒙

and the discrete number density function is given by Equation 58:

𝝋𝒊 = 𝝋(𝑿𝒊−𝟏 ) − 𝝋(𝑿𝒊 ) = ∆𝝋𝒊 Equation 58

Here the upper-case letters represent the class boundaries. The number
distribution facilitates the calculation of the average properties of the particles
either in the total population or in each size interval. This can be calculated as
follows in Equation 59:

𝑵𝑻
𝟏
̅𝑵 =
𝒙 ∑ 𝒙(𝒋) Equation 59
𝑵𝑻
𝒋=𝟏

Here 𝑥(𝑗) is the value of the characteristic for particle j and 𝑁𝑇 is the total number of
particles in the population. Grouping the particles, in this case diamonds, into

178 Stephen Coward - April 2020

 Chapter 6: Process simulation and rock characteristics

classes of particles having the same or similar characteristics allows the summation
of the characteristics of the properties of each class. It is then possible to calculate
the average property x in the whole population using Equation 60 as follows:

𝑵
𝟏
̅𝑵 =
𝒙 ∑ 𝒏(𝒊) 𝒙(𝒊) Equation 60
𝑵𝑻
𝒊=𝟏

In this case N is the number of classes or groups that have been formed, 𝑛(𝑖) is the
count of the number of particles in each group i and 𝑥(𝑖) is the characteristic defining
each group. If the groups have different sizes or densities for example it is possible
to weight this calculation by mass per group (Equation 61):

𝑵
𝟏
̅𝑵 =
𝒙 ∑ 𝒎(𝒊) 𝒙(𝒊) Equation 61
𝑴𝑻
𝒊=𝟏

where 𝑀𝑇 represents the total mass of the population being considered.

To describe the behaviour of particles within a size class a representative size is

required for each class. The objective is to define a single size that will adequately
represent the behaviour of all the particles in this class. One derivation for this
representative size is described by King (2001) as in Equation 62:

𝑫𝒊−𝟏
𝟏
𝒅𝟑𝒑𝒊 = ∫ 𝒅𝟑𝒑 𝝋(𝒅𝒑 )𝒅𝒅𝒑 Equation 62
𝝋𝒊 (𝒅𝒑 ) 𝑫𝒊

Where 𝜑(𝑑𝑝 ) is the number distribution density function, and 𝜑𝑖 (𝑑𝑝 ) is the number
fraction of the population in a size class.

For most ore streams, the sieving is conducted with a variety of screen sequences.
This process provides data where the size class width is not constant and thus a
factor needs to be used to convert the stone count frequencies or masses observed
in each class by the relative width (a measure of ‘distance’ in size space) of the
classes. For the diamond size distribution this factor, known as the unit interval, is
derived by taking the difference of the logarithms of the critical carat mass between
the upper and lower-class boundaries (Equation 63)

Stephen Coward - April 2020 179

 Evaluation of metallurgical recovery factors for diamonds recovered from kimberlites

𝟏
𝑼𝒊 = Equation 63
𝒍𝒐𝒈(𝑪𝒓𝒊𝒕𝑺𝒛𝒊 ) − 𝒍𝒐𝒈(𝑪𝒓𝒊𝒕𝑺𝒛𝒊+𝟏 )

where 𝐶𝑟𝑖𝑡𝑆𝑧𝑖 is the critical size in cts/stone of the screen on which the diamonds
have been retained, and;

𝐶𝑟𝑖𝑡𝑆𝑧𝑖+1 is the critical size in cts/stone of the screen above the retaining screen.

The critical carat mass for a given screen size is defined as that size of diamond that
will have a 50% probability of being passed or retained on a given sieve. These
factors are usually derived empirically for each diamond deposit or groups of
similar deposits. Likewise, the average representative size of the diamonds retained
on each screen size is also derived empirically. Table 26 gives the critical sizes and
average stone masses for several diamond deposits for a standard set of diamond
sized sieves.

Sieve Diamond Square Average Critical

Class Sieve Aperture Size size
Number name mm Carats Cts/stone

1 200+ 30.05 213.31 199.80

2 150+ 27.05 160.74 149.80
3 100+ 23.33 107.86 99.80
4 75+ 21.01 81.25 74.80
5 60+ 19.36 65.21 59.80
6 45+ 17.43 49.1 44.80
7 30+ 15.02 32.9 29.80
8 25+ 14.07 27.58 24.80
9 20+ 12.95 22.02 19.80
10 15+ 11.64 16.54 14.80
11 +23 8.04 8.04 7.94
12 +21 7.09 3.69 3.79
13 +19 5.56 1.92 1.95
14 +17 4.93 1.42 1.40
15 +15 4.62 1.20 1.17
16 +13 3.85 0.70 0.71
17 +12 3.42 0.52 0.51
18 +11 2.86 0.32 0.31
19 +9 2.35 0.18 0.18
20 +7 2.00 0.12 0.12
21 +6 1.72 0.08 0.08
22 +5 1.47 0.05 0.05
23 +3 1.15 0.03 0.03
24 +2 1.03 0.02 0.19
25 +1 0.82 0.01 0.01
26 -1 0.01 0.00 0.00

Table 26: Listing of Diamond screen sieve classes and associated sieve apertures, average
stone size per class and critical stone size per class.

180 Stephen Coward - April 2020

 Chapter 6: Process simulation and rock characteristics

In diamond process modelling the diamond particles are considered to have a log
normal distribution (Rombouts, 1995). This distribution can be defined as per
Equation 64:

𝒍𝒏(𝑫⁄𝑫𝟓𝟎 )
𝑷(𝑫) = 𝑮 ( ) Equation 64
𝝈

Where G(x) is the function presented in Equation 65:

𝒙
𝟏 𝟐 ⁄𝟐
𝑮(𝒙) = ∫ 𝒆−𝒕 𝒅𝒕
√𝟐𝝅 −∞ Equation 65

𝟏 𝒙
= [𝟏 + 𝒆𝒓𝒇 ( )]
𝟐 √𝟐

and where:

𝟏
𝝈= (𝒍𝒏𝑫𝟖𝟒 − 𝒍𝒏𝑫𝟏𝟔 ) Equation 66
𝟐

𝐷50 is the particle size at which P (𝐷50 ) = 0.5, this is called the median size.

The log normal distribution has theoretical significance in that it is the distribution
that results when a particle is crushed using several fracture events (Kolmogrov
1941). Its usefulness partly lies in the fact that this its properties suitably reflect the
diamond mass size distribution variable and it can be modelled using two
parameters (Aitchison and Brown, 1957). In the case of the granulometry model a
two-part fit is used. This requirement arises most likely from the incomplete
liberation of finer stones in the distribution, and or the loss and destruction of
diamonds from the coarser side of the distribution (Ferreira, 2013).

The graphic in Figure 57 shows the fitting of the curve to the recovered diamond
distribution.

Stephen Coward - April 2020 181

 Evaluation of metallurgical recovery factors for diamonds recovered from kimberlites

Actual and Modelled Diamond Size Distribution

110

100

90

80
Cumulative % Freqency Less Than Z

70

60

50

40

30

20
Actual Distribution Modelled Distribution
10

0
-3.0000 -2.0000 -1.0000 0.0000 1.0000 2.0000 3.0000

Log Lower Critical Diamond Size (log cts/stone)

Figure 57: Plot showing the actual and modelled log normal diamond size distribution.

If all the particles of kimberlite remain larger than the largest diamond, then each of
the kimberlite particles can feasibly contain a diamond of any size from the full
range of recovered diamond sizes. Once there are kimberlite particles that are
smaller than the largest diamond, the kimberlite particles can only feasibly contain
diamonds from a selected, or trimmed range of the recovered diamond size
distribution. Hence as the kimberlite ore is crushed, and the liberated diamonds
recovered, then the estimated probability of diamonds being locked in single size
fraction of discarded kimberlite is given by:

𝒊
𝑴𝒅𝒊 𝑴𝒌𝒊
𝒑𝑳 (𝑪𝒊 ) = ∑ [ ⁄∑𝒊=𝒏 𝑴𝒅 ] × [ ⁄∑𝒊=𝒏 𝑴𝒌 ] Equation 67
𝒊=𝟏 𝒊 𝒊=𝟏 𝒊
𝒊=𝟏

Where :

𝑛 is the number of size classes

𝑝𝐿 (𝐶𝑖 ) is the probability of diamonds being locked in kimberlite in size class i

𝑀𝑑𝑖 is the mass of diamond recovered in size class i

𝑀𝑘𝑖 is the mass of kimberlite in size class i

182 Stephen Coward - April 2020

 Chapter 6: Process simulation and rock characteristics

Further if one assumes the relationship shown in Equation 68:

𝒕𝒐𝒕𝒂𝒍 𝒓𝒆𝒄𝒐𝒗𝒆𝒓𝒂𝒃𝒍𝒆 𝒅𝒊𝒂𝒎𝒐𝒏𝒅 𝒄𝒐𝒏𝒕𝒆𝒏𝒕

= 𝒍𝒊𝒃𝒆𝒓𝒂𝒕𝒆𝒅 𝒅𝒊𝒂𝒎𝒐𝒏𝒅𝒔 Equation 68
+ 𝒍𝒐𝒄𝒌𝒆𝒅 𝒅𝒊𝒂𝒎𝒐𝒏𝒅𝒔

Then by rearrangement for every 100cts recovered the liberated carats can be
expressed as in Equation 69 :

𝟏𝟎𝟎
%𝑳𝒊𝒃𝒆𝒓𝒂𝒕𝒊𝒐𝒏 = 𝑿𝟏𝟎𝟎 Equation 69
𝟏𝟎𝟎 + 𝒍𝒐𝒄𝒌𝒆𝒅 𝒆𝒔𝒕𝒊𝒎𝒂𝒕𝒆𝒅

This assumption allows us to convert the locked potential calculated to an estimated

locked carat number for every 100 cts that are recovered.

The maximum diamond size that can possibly be found in a kimberlite particle that
has been discarded is, however, not only a function of the ore and diamond particle
size but must also have floated out of the dense media separation process. For this
to have happened the apparent weight of the particle must have been less than the
force of buoyancy to which the particle was exposed. A critical particle can be
defined as a particle that neither sinks nor floats, and hence the force of buoyancy
(Fb) is equivalent to the sinking force (Fs). By equating these two forces it is possible
to derive the volume and hence maximum size of diamond that can be locked in a
tailings particle for a given set of input parameters. (Equation 70)

𝐹𝑏 = 𝐹𝑠

𝑽𝑷 × 𝝆𝑴 = 𝑽𝒌 × 𝝆𝒌 + 𝑽𝒅 × 𝝆𝒅 Equation 70

Equation 71 shows the result of solving for Vd

𝑽𝒑 (𝝆𝑴 − 𝝆𝒌 )
𝑽𝒅 = Equation 71
(𝝆𝒅 − 𝝆𝒌 )

Where:

𝑉𝑑 : Maximum volume of diamond that can be contained in the particle and still float

𝑉𝑃 : Volume of particle

Stephen Coward - April 2020 183

 Evaluation of metallurgical recovery factors for diamonds recovered from kimberlites

𝜌𝑀 : The density of the cut-point that the particle is exposed to in the dense media
separation process

𝜌𝑘 : The density of the host kimberlite

𝜌𝑑 : The density of the contained diamond

Thus it is likely for most operations that the formulation in Equation 67 will be
optimistic as it assumes that a diamond of the same size as the particle can be locked
in the discarded rock. As shown in Equation 71, the probability of being locked can
be reduced to equate to cumulative proportion of the diamond size distribution that
lies below the maximum size of diamond (Dvmax) that can be locked in a kimberlite
particle of size class i. To find this class the maximum volume is converted to a mass
of diamond and then the class that this diamond of this mass would be classified into
is chosen as the "Max-i'. This leads to the probability of lockup in class i being
reformulated as in Equation 74.

𝒊=𝑴𝒂𝒙 𝒊
𝑴𝒅𝒊 𝑴𝒌𝒊
𝒑𝑳 (𝑪𝒊 ) = ∑ [ ⁄∑𝒊=𝒏 𝑴𝒅 ] × [ ⁄∑𝒊=𝒏 𝑴𝒌 ] Equation 72
𝒊=𝟏 𝒊 𝒊=𝟏 𝒊
𝒊=𝟏

This model provides the maximum size of diamond that can potentially be locked in
each size particle. Also having knowledge of the size distribution of diamonds
recovered, it is possible to calculate the possible locked distribution in each ore
particle size class and summing across all ore sizes it is then possible to construct
an estimate of the discarded diamond size distribution.

This in turn can be used to determine the maximum potential value of the discarded
stream of kimberlite particles, by multiplying the $/ct value for each size of diamond
by the estimate of locked diamond potential (see Table 33)

Application of the model

In this section the application of the granulometry model to the Venetia diamond
size frequency and a sample of two weeks production will be described.

The first part of the process is to derive the degree to which the kimberlite was
crushed, the total grind, or resulting size distribution of all the processed kimberlite
is required. This is achieved by sampling at various points in the process and
determining the relative mass flows in each stream to reconstruct the size

184 Stephen Coward - April 2020

 Chapter 6: Process simulation and rock characteristics

distribution of the processed kimberlite. The sampling collects a representative size

distribution of the mass of coarse discarded material over a two-week period and
this material is sieved into sizes. Each size fraction is then sorted into kimberlite and
non-kimberlitic particles. Once this information is available the true size
distribution can be calculated.

1 2 3 4 5 8 9 10 11
Sieve Mass % retained cumulative % kimberlite Mass Mass per size Mass Total Kimberlite
Size retained in class passing in class Kimberlite Fed to DMS Kimberlite % per class
mm (kg) % % % kg tonnes tonnes %

12.00 0.00 0.00 100.00 83.72 0.000 0 0 0.0

8.00 0.04 0.59 99.41 83.72 0.033 40 40 0.5
5.60 0.80 11.76 87.65 52.63 0.421 505 505 6.3
4.00 1.29 18.97 68.68 75.00 0.968 1160 1160 14.5
3.35 0.63 9.26 59.41 75.00 0.473 566 566 7.1
2.80 0.64 9.41 50.00 73.91 0.473 567 567 7.1
2.00 1.16 17.06 32.94 72.41 0.840 1007 1007 12.6
1.40 1.15 16.91 16.03 75.76 0.871 1044 1044 13.1
1.00 0.78 11.47 4.56 79.17 0.618 740 740 9.3
-1.00 0.31 4.56 0.00 100.00 0.310 372 2372 29.6
6.80 Totals 5.006 6000 8000 100

Table 27: Description of method used to convert the sampled tailings distribution to a total
kimberlitic distribution.

In the first column of Table 27 the aperture of the sieves used to size the sampled
material is recorded. The mass retained and the % kimberlite in each size class is
given in columns 2 and 5 respectively. This information is used to calculate the mass
of kimberlite in each size class which is shown in column 8. The size distribution is
converted to tonnes of kimberlite in each size class that will have been processed in
the period being considered by multiplying the percentage in each class by the dense
media feed tonnes for the period (6000 in this example). As there is material that is
below the bottom screen size (i.e-1mm) that will not have been sampled in the
discard material, the size distribution must be adjusted to include this material. In
this example two thousand tonnes of undersize material (also commonly referred
to as slimes) was discarded in the underflow. It is assumed that this material is all
kimberlite, which is a reasonable assumption given that most waste rock
components are usually harder than the kimberlite. This mass is added into the size
distribution in the second last row of column 10. This tonnage size distribution can
then be converted to a cumulative % retained size distribution of the processed
kimberlite for use in the model. To map the ore sizes to diamond sizes a linear
interpolation is used.

The diamonds recovered over the period are sieved and the size distribution is
modelled using a two-part log normal model.

Stephen Coward - April 2020 185

 Evaluation of metallurgical recovery factors for diamonds recovered from kimberlites

1 2 3 4 5 6 7 8 9
Diamond Critical Log Critical Actual % Fitting of Size Distribution Cumulative Error
Sieve Size Size Passing Gauss1 Gauss2 Fit Retained on model
cts/stone log cts/stone % % % % % %

200+ 199.80 2.30 100.00 100.00 100.00 100.00 0.00 0.00

150+ 149.80 2.18 100.00 100.00 100.00 100.00 0.00 0.00
100+ 99.80 2.00 100.00 100.00 100.00 100.00 0.00 0.00
75+ 74.80 1.87 99.99 100.00 100.00 100.00 0.00 0.00
60+ 59.80 1.78 99.96 100.00 100.00 100.00 0.00 0.00
45+ 44.80 1.65 99.92 100.00 100.00 100.00 0.00 0.01
30+ 29.80 1.47 99.80 99.99 100.00 99.99 0.01 0.04
25+ 24.80 1.39 99.72 99.99 100.00 99.99 0.01 0.07
20+ 19.80 1.30 99.60 99.97 100.00 99.97 0.03 0.14
15+ 14.80 1.17 99.38 99.94 100.00 99.94 0.06 0.31
+23 8.04 0.91 98.71 99.77 100.00 99.77 0.23 1.11
+21 3.69 0.57 97.38 98.85 100.00 98.85 1.15 2.18
+19 1.92 0.28 94.61 96.47 100.00 96.47 3.53 3.47
+17 1.42 0.15 92.58 94.47 99.98 94.47 5.53 3.56
+15 1.20 0.08 91.10 92.94 99.96 92.94 7.06 3.38
+13 0.70 -0.15 84.97 86.29 99.66 86.29 13.71 1.75
+12 0.52 -0.28 80.40 81.13 99.03 81.13 18.87 0.54
+11 0.32 -0.50 70.89 70.07 95.70 70.07 29.93 0.68
+9 0.18 -0.75 57.12 54.75 84.33 54.75 45.25 5.61
+7 0.12 -0.93 44.11 42.73 68.46 42.73 57.27 1.92
+6 0.08 -1.10 32.29 32.24 49.85 32.24 67.76 0.00
+5 0.05 -1.31 17.60 20.89 27.02 20.89 79.11 10.80
+3 0.03 -1.59 6.99 10.29 8.00 8.00 92.00 1.03
+2 0.02 -1.73 4.06 6.78 3.58 3.58 96.42 0.23
+1 0.01 -1.97 1.76 2.92 0.62 0.62 99.38 1.30
-1 0.01 -2.30 0.00 0.76 0.03 0.03 99.97 0.00

Table 28: Fitting of the lognormal model to the recovered size distribution.

Table 28 shows the resulting fit and the squared error in each of the size classes.
The graph of the actual and fitted distribution is shown in Figure 57. Although some
classes show some deviation between the actual and the model fit, in general the
form of the model and actual recovery correspond reasonably well.

The parameters of the process that has been used to separate the kimberlite is also
recorded and is given in Table 29.

For each composite particle size, Equation 60 can be used to calculate the maximum
size of diamond that can be locked in a ‘critical’ class particle. Once this has been
done for each size class considered it is possible to calculate what proportion of the
diamond distribution can be locked in this particle.

186 Stephen Coward - April 2020

 Chapter 6: Process simulation and rock characteristics

Plant Operating Parameters

Diamond Density 3.515 g/cm3
Kimberlite Density 2.5 g/cm3
Operating density 3 g/cm3
Actual Cut Point 2.7 g/cm3
Headfeed 8000 dry tonnes per day
Tonnes to oversize
0 dry tonnes per day
discard
Tonnes feed to fines DMS 6000 dry tonnes per day
Tonnes to undersize 2000 dry tonnes per day
Table 29: Process parameters required for the granulometry model.

This diamond mass will most likely have a distribution that is equivalent to the
recovered distribution truncated by the maximum locked diamond particle size. The
result of these calculations is shown in Table 30.

Sieve Parameters Kimberlitic Breakdown Kimberlite Distribution Particle Size Maximum Locked diamond
Diamond Square Average Critical Cum % Class log Square Spherical Factored vol Volume Size Size
Sieve mesh Size size tonnes tonnes Mesh Volume 0.9 diamond carats mm
name mm cts/stone cts/stone % % log mm cm3 cm3 cm3 cts mm
32 0.00 1.51 17.16 15.44 2.45 43.04 18.58
200+ 30.05 213.31 199.80 0.00 0.00 1.48 14.21 12.79 2.03 35.64 17.45
150+ 27.05 160.74 149.80 0.00 0.00 1.43 10.36 9.33 1.48 26.00 15.71
100+ 23.33 107.86 99.80 0.00 0.00 1.37 6.65 5.98 0.95 16.68 13.55
75+ 21.01 81.25 74.80 0.00 0.00 1.32 4.86 4.37 0.69 12.18 12.20
60+ 19.36 65.21 59.80 0.00 0.00 1.29 3.80 3.42 0.54 9.53 11.24
45+ 17.43 49.1 44.80 0.00 0.00 1.24 2.77 2.50 0.40 6.96 10.12
30+ 15.02 32.9 29.80 0.00 0.00 1.18 1.77 1.60 0.25 4.45 8.72
25+ 14.07 27.58 24.80 0.00 0.00 1.15 1.46 1.31 0.21 3.66 8.17
20+ 12.95 22.02 19.80 0.00 0.00 1.11 1.14 1.02 0.16 2.85 7.52
15+ 11.64 16.54 14.80 0.05 0.05 1.07 0.83 0.74 0.12 2.07 6.76
+23 9.28 8.04 7.94 0.34 0.30 0.97 0.42 0.38 0.06 1.05 5.39
+21 7.09 3.69 3.79 2.85 2.51 0.85 0.19 0.17 0.03 0.47 4.12
+19 5.56 1.92 1.95 6.81 3.96 0.75 0.09 0.08 0.01 0.23 3.23
+17 4.93 1.42 1.40 12.66 5.85 0.69 0.06 0.06 0.01 0.16 2.86
+15 4.62 1.20 1.17 15.54 2.88 0.66 0.05 0.05 0.01 0.13 2.68
+13 3.85 0.70 0.71 22.94 7.39 0.59 0.03 0.03 0.00 0.07 2.24
+12 3.42 0.52 0.51 27.62 4.68 0.53 0.02 0.02 0.00 0.05 1.99
+11 2.86 0.32 0.31 34.70 7.08 0.46 0.01 0.01 0.00 0.03 1.66
+9 2.35 0.18 0.18 44.95 10.25 0.37 0.01 0.01 0.00 0.02 1.36
+7 2.00 0.12 0.12 48.05 3.11 0.30 0.00 0.00 0.00 0.01 1.16
+6 1.72 0.08 0.08 54.14 6.09 0.24 0.00 0.00 0.00 0.01 1.00
+5 1.47 0.05 0.05 59.58 5.44 0.17 0.00 0.00 0.00 0.00 0.85
+3 1.15 0.03 0.03 66.89 7.30 0.06 0.00 0.00 0.00 0.00 0.67
+2 1.03 0.02 0.19 69.66 2.78 0.01 0.00 0.00 0.00 0.00 0.60
+1 0.82 0.01 0.01 75.70 6.04 -0.09 0.00 0.00 0.00 0.00 0.48
-1 0.01 0.00 0.00 100.00 24.30 -2.00 0.00 0.00 0.00 0.00 0.01
100.00

Table 30: Calculation of the maximum locked diamond size in kimberlite particles in each
sieve class.

While particles are assumed to have a spherical shape, their volume, expressed in
cm3, is based on the square mesh size, 𝑑𝑖 , specified in mm (Equation 73).

Stephen Coward - April 2020 187

 Evaluation of metallurgical recovery factors for diamonds recovered from kimberlites

𝟒 𝒅𝒊 𝟑
𝑽𝒊 = × 𝝅 × ( ) × 𝟏𝟎𝟎𝟎 Equation 73
𝟑 𝟐

This volume must be factorised to some volume lower than a sphere. In the model
presented a value of 0.9 is used. This value can be calibrated for various ore types
and crushing methods, by direct measurement of the volume of particle volumes in
each sieve class. Once the representative volume of each sieve class has been
calculated it is possible, using the parameters given in Table 29 and Equation 71, to
calculate the maximum volume (in cm3) and mass of diamond that can be locked up
in this particle size class.

It is now possible to work out the proportion of the size distribution of diamonds
that can be locked in each size class. This calculation is demonstrated in Table 32.
The parameters used to fit a log normal distribution to the recovered diamond size
distribution (Table 28) are used to determine this percentage given the maximum
locked diamond size. However, this percentage of the distribution represents a
number of carats locked for every 100 carats recovered and will most likely occur
in the same size distribution as that of the recovered diamonds. The single rider is
that if the carats locked in any given size is smaller than the lower critical size of that
sieve class, the locked carat potential is moved down to the next smallest size class.
This calculation is show in Table 31.

Sieve Parameters Allocation of Locked potential in each size class using the recovered diamond distribution
Diamond Square Reallocated
Sieve mesh Cts/class
name mm +19 +17 +15 +13 +12 +11 +9 +7 +6 +5 +3 +2 +1 -1 cts
32
+19 5.56 0.13 0.13
+17 4.93 0.09 0.06 0.16
+15 4.62 0.07 0.05 0.02 0.14
+13 3.85 0.28 0.19 0.09 0.15 0.71
+12 3.42 0.21 0.14 0.06 0.11 0.06 0.59
+11 2.86 0.43 0.30 0.13 0.24 0.12 0.10 1.32
+9 2.35 0.62 0.44 0.20 0.34 0.17 0.14 0.11 2.02
+7 2.00 0.59 0.41 0.18 0.33 0.16 0.14 0.10 0.00 1.91
+6 1.72 0.53 0.37 0.17 0.30 0.15 0.12 0.09 0.00 0.00 1.74
+5 1.47 0.66 0.46 0.21 0.37 0.18 0.15 0.12 0.00 0.00 0.00 2.16
+3 1.15 0.48 0.34 0.15 0.27 0.13 0.11 0.08 0.00 0.00 0.00 0.03 1.59
+2 1.03 0.13 0.09 0.04 0.07 0.04 0.03 0.02 0.00 0.00 0.00 0.01 0.01 0.44
+1 0.82 0.10 0.07 0.03 0.06 0.03 0.02 0.02 0.00 0.00 0.00 0.01 0.01 0.00 0.35
-1 0.01 0.08 0.06 0.03 0.04 0.02 0.02 0.01 0.00 0.00 0.00 0.00 0.00 0.00 0.01 0.27
Total locked per class 4.40 2.99 1.31 2.28 1.06 0.84 0.56 0.00 0.00 0.00 0.04 0.02 0.00 0.01 13.52

Table 31: Allocation of the locked potential in each size class according to the probabilities
derived from the recovered diamond size distribution.

The re-assigned locked potential is shown in Table 32 in the column titled ‘Re-
distributed per class cts’.

188 Stephen Coward - April 2020

 Chapter 6: Process simulation and rock characteristics

Sieve Parameters Determination of Locked Diamond Distribution Total Estimated Distribution

Diamond Square % of Diam. max locked Re-Distributed % per Clss Locked Recovered Total
Sieve mesh Distribution Cts in class per class Locked Stones Distribution distribtuion
name mm < crit size cts/100 rec cts % number cts cts
32 100.00
200+ 30.05 99.99 0.00 0.0000 0.0000 0.00 0.00 0.00
150+ 27.05 99.99 0.00 0.00 0.00 0.00 0.00 0.00
100+ 23.33 99.96 0.00 0.00 0.00 0.00 0.00 0.00
75+ 21.01 99.91 0.00 0.00 0.00 0.00 0.01 0.01
60+ 19.36 99.84 0.00 0.00 0.00 0.00 0.03 0.03
45+ 17.43 99.68 0.00 0.00 0.00 0.00 0.04 0.05
30+ 15.02 99.20 0.00 0.00 0.00 0.00 0.12 0.14
25+ 14.07 98.83 0.00 0.00 0.00 0.00 0.08 0.09
20+ 12.95 98.17 0.00 0.00 0.00 0.00 0.12 0.14
15+ 11.64 96.87 0.04 0.00 0.00 0.00 0.21 0.24
+23 9.28 91.60 0.27 0.00 0.00 0.00 0.67 0.76
+21 7.09 78.93 1.98 0.00 0.00 0.00 1.34 1.52
+19 5.56 61.20 2.42 0.13 0.93 0.07 2.77 3.14
+17 4.93 51.11 2.99 0.16 1.15 0.11 2.03 2.30
+15 4.62 45.58 1.31 0.14 1.00 0.11 1.48 1.69
+13 3.85 30.84 2.28 0.71 5.26 1.01 6.13 6.96
+12 3.42 22.56 1.06 0.59 4.34 1.12 4.57 5.19
+11 2.86 11.93 0.84 1.32 9.77 4.17 9.51 10.79
+9 2.35 2.81 0.29 2.02 14.96 11.30 13.77 15.63
+7 2.00 0.60 0.02 1.91 14.14 16.33 13.01 14.77
+6 1.72 0.11 0.01 1.74 12.84 21.92 11.82 13.42
+5 1.47 0.01 0.00 2.16 15.96 44.49 14.69 16.68
+3 1.15 0.00 0.00 1.59 11.73 61.95 10.62 12.05
+2 1.03 0.00 0.00 0.44 3.29 23.92 2.93 3.32
+1 0.82 0.00 0.00 0.35 2.58 33.05 2.30 2.61
-1 0.01 0.00 0.00 0.27 2.03 108.09 1.76 2.00
13.52 13.52 100.00 327.64 100.00 113.52

Table 32: Calculation of the locked carat potential per diamond sieve class.

In the example presented here approximately 13.5 carats are locked for every 100
hundred recovered. The liberation derived for this period would be calculated as
per Equation 69, and result in estimated carat liberation of 88%. Knowing the
recovered size distribution and the locked potential facilitates the calculation of the
locked and liberated revenue. This is essentially a weighting of the distribution by
the revenue obtained for each size of diamond. This calculation is summarised in
Table 33.

Stephen Coward - April 2020 189

 Evaluation of metallurgical recovery factors for diamonds recovered from kimberlites

Sieve Parameters Recovered locked Estimated Total Distribution Revenue Distribution Caclulation Cumulative% of Revenue
Diamond Square Carats per Redistributed Total Total Cumulative Revenue Revenue Revenue
Sieve mesh Class locked Contained Contained % retained per class Free Locked Free Locked
name mm Cts Cts Cts per class % per class % $/ct $/100cts $/100cts % %
31.5
200+ 30.05 0.00 0.00 0.00 0.00 0.00 100.00 0.00 0.00 0.00 0.00
150+ 27.05 0.00 0.00 0.00 0.00 0.00 100.00 0.00 0.00 0.00 0.00
100+ 23.33 0.00 0.00 0.00 0.00 0.00 100.00 0.00 0.00 0.00 0.00
75+ 21.01 0.01 0.00 0.01 0.01 0.01 100.00 1.01 0.00 0.06 0.00
60+ 19.36 0.03 0.00 0.03 0.02 0.03 100.00 2.65 0.00 0.20 0.00
45+ 17.43 0.04 0.00 0.04 0.04 0.07 100.00 4.03 0.00 0.43 0.00
30+ 15.02 0.12 0.00 0.12 0.11 0.18 100.00 12.41 0.00 1.12 0.00
25+ 14.07 0.08 0.00 0.08 0.07 0.25 100.00 7.76 0.00 1.55 0.00
20+ 12.95 0.12 0.00 0.12 0.11 0.35 100.00 12.31 0.00 2.24 0.00
15+ 11.64 0.21 0.00 0.21 0.19 0.54 100.00 21.40 0.00 3.44 0.00
+23 9.28 0.67 0.00 0.67 0.59 1.13 75.00 50.43 0.00 6.25 0.00
+21 7.09 1.34 0.00 1.34 1.18 2.31 70.00 93.53 0.00 11.47 0.00
+19 5.56 2.77 0.13 2.89 2.55 4.86 65.00 179.94 8.14 21.51 0.45
+17 4.93 2.03 0.16 2.18 1.92 6.78 40.00 81.02 6.23 26.03 0.80
+15 4.62 1.48 0.14 1.62 1.43 8.21 35.00 51.96 4.73 28.93 1.07
+13 3.85 6.13 0.71 6.84 6.03 14.24 30.00 183.94 21.35 39.19 2.26
+12 3.42 4.57 0.59 5.16 4.54 18.78 26.00 118.78 15.26 45.82 3.11
+11 2.86 9.51 1.32 10.83 9.54 28.32 15.00 142.59 19.82 53.77 4.21
+9 2.35 13.77 2.02 15.79 13.91 42.23 12.00 165.23 24.27 62.99 5.57
+7 2.00 13.01 1.91 14.92 13.14 55.37 11.00 143.10 21.02 70.97 6.74
+6 1.72 11.82 1.74 13.55 11.94 67.31 10.00 118.19 17.36 77.57 7.71
+5 1.47 14.69 2.16 16.85 14.84 82.15 8.00 117.53 17.26 84.13 8.67
+3 1.15 10.62 1.59 12.20 10.75 92.90 7.00 74.32 11.10 88.27 9.29
+2 1.03 2.93 0.44 3.37 2.97 95.87 6.00 17.55 2.67 89.25 9.44
+1 0.82 2.30 0.35 2.65 2.33 98.21 5.00 11.49 1.75 89.89 9.54
-1 0.01 1.76 0.27 2.04 1.79 100.00 5.00 8.81 1.37 90.39 9.61
TOTAL 100.00 13.52 113.52 100.00 1420.00 1620.00 172.33

Table 33: Calculation of the liberated and locked potential revenue.

Using the same form of calculation as in Equation 74 we can derive an estimate of

the liberated revenue as follows:

% 𝑹𝒆𝒗𝒆𝒏𝒖𝒆 𝑳𝒊𝒃𝒆𝒓𝒂𝒕𝒊𝒐𝒏
𝟏𝟎𝟎 Equation 74
= 𝑿𝟏𝟎𝟎
𝟏𝟎𝟎 + 𝒍𝒐𝒄𝒌𝒆𝒅 𝒓𝒆𝒗𝒆𝒏𝒖𝒆

In this case the revenue liberation would be estimated to be in the vicinity of 90%,
suggesting that there remains an additional 11% of current revenue to be recovered
if the crushing circuit was modified to release all of the contained diamonds.

Using this approach to determine liberation allows for the identification of which
size fractions are most important in terms of locking up carats and which will have
the largest impact on the variation in liberated revenue. In this way it is possible to
carry out trade-off studies to determine steps to find the best balance between
process cost and revenue.

190 Stephen Coward - April 2020

 Chapter 6: Process simulation and rock characteristics

Generating a flowsheet using calibrated process models

It is possible, using a platform such as Microsoft Excel, to combine the unit process
models to create a model of a diamond liberation and dense media separation plant.
An example is provided in the appendix and has been created using the add-in
package “Limn” developed, and kindly sponsored for this research, by Dave
Wiseman. Using this model, it is possible to determine the total grind, estimate DMS
separation efficiency and estimate the recovery and loss of diamonds that will be
achieved for each given set of rock properties when the plant is in a stable state.

The output of this model can be stored in a matrix that can be used in an integrated
way to interact with the block model containing spatially varying rock properties.

Use of the liberation model to derive a metallurgical recovery factor

In cases where this model is shown to be a robust estimator of the diamonds that
are liberated and locked for a given size distribution of ore particles, the process
simulation that is required will consist of a comminution model that produces an
estimate of the final grind size distribution. The comminution model can be
integrated with a spatial model of the properties of the rocks that govern the final
grind, then a methodology can be devised that will facilitate the relationship
between ore properties and diamond recovery.

6.6 Process simulation and sensitivity testing

Dynamic process simulation modelling introduces a time dimension to the

processes of modelling diamond recovery. Using this approach, it is possible to
stress the various settings in the process model and to determine how the steady
state operation will change if several of the parameters of the process can oscillate
through a realistic range with some predetermined probability distribution of that
range.

6.7 Limitations of processing models

There are several limits to both the granulometry model and the process
simulations model and their use in integrated models. These include:

• Difficulty of defining the relationships between individual rock properties

and the impact of rock mixes on unit process models;

Stephen Coward - April 2020 191

 Evaluation of metallurgical recovery factors for diamonds recovered from kimberlites

• Stress testing the process responses when the material changes and the
process responds in different ways; and
• Time that the process takes to reach steady state and its incorporation into
process models.

Each of these difficulties has implications for the reliability of the estimation of the
recovery factor these will be explained in some detail in this section.

6.8 Conclusion

It is possible to estimate the expected total grind that will be achieved for a given
process flowsheet and ore type characteristic given sufficient data. Ideally for
existing processes the experiments can be designed to define relationships between
operating parameters, ore types and the resulting diamond recovery. For new
deposits without this information it is possible to generate some understanding of
the range of expected recoveries and to determine which processes are likely to
constrain impact on recovery.

Process models can be used in integrated evaluation processes provided a matrix of

inputs, parameters and outputs can be created for the range of rock types that are
to be expected. This does, however, mean that it is very difficult to test the sensitivity
of diamond recovery to changes in process parameters or uncertainty in rock
properties as there is a distinctly non-linear relationship between changes to rock
types and the outcomes of the process.

In the subsequent chapters the methods developed here are integrated with
orebody sampling, simulation and estimation approaches suggested in the prior
chapters to develop an integrated value chain model that simulates the interaction
of variable ore characteristics with the mining and processing models to produce
period by period production outputs.

192 Stephen Coward - April 2020

 Chapter 7: Value chain models - case study 1

7 VALUE CHAIN MODELS -

CASE STUDY 1

7.1 Introduction

This case study demonstrates an integrated value chain model that comprises
multiple orebody realisations with mining and treatment process models. The value
chain model is used to simulate the mine operation and generate daily production
summaries. This output is used to quantify the impact that the rock characteristics
have on the range and variability of the metallurgical recovery factor.

The case study was published as “Integrated Mine Evaluation – Implications for
Mine Management” in the proceedings of the AusIMM Mine Managers Conference
(Nicholas et al., 2007). The author of this thesis presented this paper at this
conference held in Melbourne, Australia in 2007, and was responsible for
developing the orebody models, the VBA coding, execution of mining and process
simulations and providing links into the financial model. The work was completed
in collaboration with Grant Nicholas, Kurt Petersen, Alain Galli and Margaret
Armstrong.

In this case study the in situ diamond grade and revenue were relatively well
constrained. However uncertain orebody geometry and the impact that the complex
geometry would have on mining rate and recovery, were not.

The case study demonstrates how it is possible to acquire data from the orebody to
spatially simulate the vertical displacement, the thickness of the dyke and the grade
using conventional geostatistical techniques. These characteristics of the orebody
are used in a mining simulation to generate a variable mining rate and feed
characteristics for the process plant. The process plant model was configured to
respond to the feed ore characteristics and predict throughput, comminution and
diamond recovery.

Using the novel approach developed during this research it was possible to model
and simulate the key aspects of this project including:

Stephen Coward - April 2020 193

 Evaluation of metallurgical recovery factors for diamonds recovered from kimberlites

• Generating a synthetic version of the orebody that was simulated at a small

block size that resembled the orebody geometry, variable thickness and
grade;

• Developing a method to sample the synthetic orebody at different drill hole

spacings to generate multiple estimates and simulations of the orebody and
hence incorporate sampling uncertainty in the model;

• Developing a method to assign a feasible block sequence to each ore block

used in a mining simulation;

• Simulating mining that adheres to the planned sequence of mining with the
rate of mining each block being determined by the interaction of the
morphology of the dyke and mining equipment used;

• Determination of dilution incurred when mining blocks as a function of dyke

thickness, planned mining height and vertical displacement of dyke location;
and

• Using dilution as an input into the process model to calculate throughput,

liberation and recovery.

The simulated production outputs were used as inputs into a financial model to
calculate the value of each of the alternatives considered, defined by the Net Present
Value. The benefit of using a “Master Synthetic Orebody model” or so-called “V-Bod”
model to represent reality is that it is possible to compare and then track the
differences between the processing of the Master model (simulated reality at small
scale) and the models generated from the sampling of the Master model and using
these sample values to generate estimated resource models.

Project Background

The orebody is a kimberlite dyke that dips at 15 degrees towards the north east and
is on average 2.8m thick. Most of the kimberlite dyke is hosted within an Archean
multiphase suite of intrusive granitoids, with a minor portion of the kimberlite dyke
emplaced within overlying metavolcanics and metasediments of the Archean
greenstone belts.

The geometry of the dyke is variable. On the regional scale (100’s of metres) the
Snap Lake dyke appears to be a ‘simple’ continuous, gently dipping sheet, although
three areas of offset have been identified by surface seismic imaging' (McBean et al.,
2001). At a more local scale (10-100m), orientation changes and splits and large

194 Stephen Coward - April 2020

 Chapter 7: Value chain models - case study 1

splays have been observed. These are thought to be structurally controlled. On a

small scale (0-10m), the dyke is typically controlled by two different host rock
features. Within the strongly foliated metavolcanics, the kimberlite appears to roll
and undulate on a small scale matching the foliation, while in the granitic host rocks,
local variations occur along a primary set of joints that are flat lying but affected by
secondary jointing at approximately 35° dip resulting in an angular step-like nature
of the dyke (McBean et al., 2001).

This variability in the geometry of the deposit has several implications for grade
estimation, mine design and plant operation, including but not limited to:

• Uncertainty in the volume of the dyke volume in each block;

• The rate at which development and mining can be achieved;
• Dilution that will be incurred when mining;
• Waste ore mix that will be fed to the process; and
• Impact of dilution on recovery efficiency.

The mine plan was based on kriged estimates of the grades, volumes and other
variables and is assumed, on average, to provide a way to develop an unbiased
prediction of the annual production. In the derivation of the plan no allowance was
made for the short impact of short-scale variability (in both the spatial and temporal
sense) on the rate and efficiency (ore recovery, ore loss, diamond recovery) of the
mining and treatment processes.

Orebody model data and model preparation

Sampling data that is used in any evaluation play a fundamental part in producing
estimates that aim to reflect an unknowable reality. Although the inclusion of more
samples reduces uncertainty associated with both the mean and variance of
resource estimates, it does not alter the true "natural" variability within the deposit.
It does however alter out quantification of it which can only be estimated from the
results of the samples.

The limitations of designing a sampling campaign for multiple variables have been
discussed by Kleingeld and Nicholas (2004). In this case study three orebody
variables were considered in the evaluation model:

• The geometrical variability of the top surface of the dyke (v1);

• Thickness, which can be used to determine the volume of the dyke; and
• Grade (in carats per hundred tonnes)

Stephen Coward - April 2020 195

 Evaluation of metallurgical recovery factors for diamonds recovered from kimberlites

Synthetic core drilling was used to delineate geological variability on three different
grid densities; 50 by 50m, 25m by 25m and 10m by 10m, creating scenarios one,
two and three. These were designed to sample within the expected variogram
ranges of 75m for thickness and 50m for v1. Sampling campaigns set beyond the
range would return the variable’s average value and would not detect the short-
scale variability. A 50m by 50m drilling grid was used to sample for grade, using
large diameter drilling (LDD). Grade was not deemed to have any significant
variability between scenarios and therefore, a single sampling campaign was
deemed sufficient.

Table 1 describes the design of the simulated sampling campaigns on the virtual
orebody (V-bod); sampling occurred at point support and simulation grid nodes
were 4m by 4m in dimension.

V-bod Scenario 1 Scenario 2 Scenario 3

Wide Moderate
Assumed Close Spaced
Description spaced spaced
Reality Samples
samples samples
Grid Dimensions 4m x 4m 50m x 50m 25m x 25m 10m x 10m
No of Samples 45 000 288 1152 7200
Samples as a % of the V-bod Blocks - 0.64 2.56 16

Table 34: Summary of the characteristics of three sampling campaigns and that used to
define the 'Virtual Orebody(V-bod )

The V-bod was created using a non-conditional geostatistical simulation based on

data gathered from a combination of drilling information, bulk-samples and face
mapping from an exposed part of the dyke. It is assumed to be the ‘reality’ on which
the various sampling campaigns were simulated to generate virtual sample data.
These sample data were used to generate kriged estimates and resource simulations
of grade, dyke thickness and geometric surface undulations of the dyke. A mine plan
was overlain onto each estimate and the complete simulation to determine the
reserves. All production output was fed into the financial model.

The graphic base maps of the V-bod and each sampling campaign are shown in
Figure 58 (warmer colours represent higher values while darker colours are low
values).

196 Stephen Coward - April 2020

 Chapter 7: Value chain models - case study 1

Virtual Ore Body

V1 Thickness Grade

Scenario 1

V1 Thickness Grade

Scenario 2

V1 Thickness Grade

Scenario 3

V1 Thickness Grade

Figure 58: Comparison of the thickness and v1 base maps for the kriged and simulated
outputs of each scenario with that of the V-bod. Grade was held constant for each scenario.

Stephen Coward - April 2020 197

 Evaluation of metallurgical recovery factors for diamonds recovered from kimberlites

Table 2 - Resource Simulation Output

V-bod Scenario 1 Scenario 2 Scenario 3
Mean Thickness 1.67 1.62 1.64 1.65
Std dev Thickness 0.43 0.52 0.45 0.45
Mean v1 1.88 1.79 1.85 1.87
Std dev v1 0.34 0.40 0.39 0.35
Mean Grade 196.50 197.30 196.10 197.00
Std dev Grade 35.51 38.08 34.41 35.43

Table 35: The descriptive statistics for the V-bod and each scenario for grade, dyke thickness
and the geometrical variability of the dyke surface (v1).

Processing system model

The block model data for each realisation (the kriged model is one realisation, and
the 25 simulated realisations constituted the remainder) were exported from the
Isatis software into a text file containing the information associated with each block
in individual rows. These files can then be processed individually through the
system. The model is built in an Excel VBA framework, and consists of several
modules that represent the functions for mining, stockpiling, crushing and dense
media separation processes.

Mining parameters and constraints

The mine simulation presented an opportunity to evaluate the impact of the dyke
geometry on mining selectivity and rate.

The proposed mining method utilised a mixed fleet of trackless vehicles to mine in
a stope and pillar configuration. The designed mine plan targeted an average
extraction rate of 75 percent of in situ kimberlite at average daily delivery of 3150
tonnes of kimberlite with minimal dilution . Each mining panel, approximately 250
m by 250 m was mined in a sequence that required the establishment of rim tunnels,
then stope tunnels and finally excavations to facilitate stope slashing or drifting.

The model operated as the scale of the smallest mining unit (SMU) with each block
having planar dimensions of 4 m by 4 m, but with variable heights. The rate, and
hence time, to mine each SMU was determined at run time as a function of the
machine used to carry out the mining and the height that was mined. The mined
height, (and hence mined volume, tonnage, dilution and grade) was determined by
the tunnel type, a set of mining height constraints for each tunnel type were selected
as follows :

• rim tunnels maximum of 3.5 m (height)

198 Stephen Coward - April 2020

 Chapter 7: Value chain models - case study 1

• stope tunnels minimum 2.0 m (height), and

• stope blocks minimum 1.0 m (height), for small machines and 1.5m (height) for
large machines.

Pillar dimensions varied depending on the support required as a function of fracture

frequency/ground conditions but a maximum span limit of 8 m was imposed in all
tunnel configurations. The change in vertical location of the footwall in rim and
stope tunnels was constrained to an angle that would not exceed the safe
incline/decline on which the mining machines could operate. This constraint lead
to loss of kimberlite in the hanging wall in some cases. The implementation of these
constraints is depicted in the diagram in Figure 59.

Figure 59: A diagram depicting the implementation of the mining constraint logic.

Processing parameters

The system has several constraints that can be modified. A set of constraints that is
used in processing was referred to in this case as a scenario. Note that in later
models, and publications, this term has been specifically reserved for future
operating contexts.

Stephen Coward - April 2020 199

 Evaluation of metallurgical recovery factors for diamonds recovered from kimberlites

When processing the orebody models through the value chain the variability of the
orebody geometry impacts on the mining rate and the relative proportion of
kimberlite and waste rock mined. This mix of kimberlite and waste contained in the
material processed impacts on the plant throughput rate, the degree of
comminution and the recovery efficiency of the DMS.

To determine the relationship between the dilution and process efficiency, data
were acquired from the sample plant that had been treating samples from the
deposit. A process simulation model was developed using Limn software in
collaboration with Dr. K. Petersen (Petersen, 2005) . The simulation represented the
planned full-scale process plant using the sample plant data to calibrate both
expected size reduction and recovery. The model was run for several different
combinations of granitoid waste, metavolcanic waste and kimberlite. The resulting
data was used to investigate the relationship between the proportion of dilution and
comminution and DMS performance. It was observed that the comminution and
DMS responses could be simplified and represented by a least squares linear fit to
the simulation outputs. This approach allow content range of responses for differing
proportions of waste and kimberlite vs diamond recovery were plotted and a least
squares linear model was fitted.

Diamond Liberation Vs. Kimberlite in Headfeed

Granitic Dilution
100

95

90

85
% Diamond Liberation

80

75

70

65

60

55

50
30 40 50 60 70 80 90 100
% Kimberlite in Headfeed(%)

RawTestData Reference Case Fitted Parameters

Lower Case Upper Case
Simulated Curves

Figure 60: A depiction of liberation as a function of the concentration of kimberlite in the

headfeed.

Once relationship was established it was possible to use the standard error of the
regression line to determine the uncertainty in the parameters of the fit and use this

200 Stephen Coward - April 2020

 Chapter 7: Value chain models - case study 1

to seed a simulation of the value for these parameters (Figure 60). The parameters,
in this case M and C for the regression line can be randomly drawn from a calibrated
normal distribution to generate a set that can be used in the value chain model.
Simulated parameters that produce a line that falls outside the limits found at other
similar operations are rejected from the set of feasible parameters. In this way it is
possible to introduce a range of uncertainty that is constrained by prior observation
and gives a realistic model to the variability in recovery that the changing ore
properties will drive.

The use of these simplified functions facilitates simulation of the entire life of mine
at a block scale and to reflect the interactions of the ore characteristics with the
process. This would not currently be possible in realistic timeframes suing a full
population balance model.

Interpretation and use of outputs from the value chain model

The processing of the mineral resource block models generates comprehensive

production data that are evaluated using a comprehensive financial model. This
approach allowed technical range analysis to be converted into meaningful and
comparative financial terms. The outputs can be used to rank and thus identify the
mining and processing constraints that impact most on the range of output.
Consequently, the value chain model can be used iteratively in a structured way to
search for so-called ‘plateaus of value’ where the trade-off between an overall
maximum value (using a financial indicator such as NPV) and a value that is
relatively robust in the face of uncertain operational outcomes is revealed.

7.2 Benefits and limitations of the methodology

The model as conceived and represented in this research paper demonstrates an

implementation of the methodology that can be used to define confidence levels
around the expected metallurgical recovery factor and the project value for various
configurations. The validity of the findings requires that several assumptions about
the uncertainty in the orebody and the interaction of this uncertainty with
constraints in the value chain hold true for the duration of the operation. These
assumptions and the implications they have for the design and management of this
mine are briefly discussed below.

The model can be used to demonstrate how changes in dyke footwall location will
impact on both development and mining rates. When there were changes in footwall
location that resulted in mining roadway slopes that are too steep to be safely used
by the mining equipment, the mining teams would have to retreat and redevelop
ramps and roadways. The additional activity would reduce production output. The

Stephen Coward - April 2020 201

 Evaluation of metallurgical recovery factors for diamonds recovered from kimberlites

model also shows how it is possible to evaluate impact of this constraint in areas of
high dyke variability in which small machines with less tolerance for steep
roadways are typically allocated. This unexpected result suggests that more cover
drilling may be required to inform roadway design and machine allocation planning.

The model can be used to show how the impact of reduced mining rates can to some
extent be managed by additional advance development to create flexibility to mine
ore from several operating areas simultaneously. It can be shown that adopting this
strategy will increase the dilution of ore fed to the process plant whilst setting up
panels to mine, but the average dilution in the later years of operation would be able
to be brought inline with initial expectations.

Establishing a relationship between the rock mix fed to the plant, the comminution,
liberation and hence recovery, required a combination of sample plant processing
and process modelling to predict the nature of the relationship. The assumption that
the impact of diluting waste on comminution that was observed at sample plant
scale (Samples comprising ~80 tonnes each) will be observed at production scale
appears reasonable but will have to be verified during commissioning of the full-
scale plant. This relationship does, however, highlight that evaluation of the impact
of dilution without using a block-by-block evaluation model (as opposed for
instance with doing a sensitivity analysis in the financial model) will most likely
underestimate the impact that dilution will have on the operation.

This model has assumed that the 3000-tonne storage to buffer surges between the
mine and the plant is a fixed maximum. The model can be used to demonstrate how
increasing this capacity constraint will have a material impact on the value of the
operation. The flexibility that additional surge capacity will provide can be valued
and hence the capital that is allocated to this part of the operation can be justified.
If a smoothed model of the orebody, or a smoothed feed rate was assumed the value
of storage would be materially underestimated.

The value chain modelling approach can accommodate not only unsystematic risk
(risks associated with the specifics of the envisaged project) but can also be used
simultaneously to evaluate the impact of systematic risk, i.e. risks that are
independent of the project configuration, such as exchange rate and diamond sales
price. In this holistic approach it is possible to contrast and compare the relative
merits of technical risk mitigation and mitigating project risk through various
financial engineering measures.

202 Stephen Coward - April 2020

 Chapter 8: Value chain model - case study 2

8 VALUE CHAIN MODEL - CASE

STUDY 2

8.1 Introduction

This case study demonstrates an integrated value chain model that comprises
multiple orebody realisations with several alternative mining and treatment
process options. The value chain model is used to simulate the mine operation and
generate daily production summaries. This output is used to quantify the impact
that the rock characteristics have on the range and variability of the metallurgical
recovery factor.

The main source of uncertainty in this project was the impact that hard and dense
kimberlite breccias would have on diamond recovery and process performance. The
hardness of the kimberlite was expected to compromise liberation, and the zones of
higher density were expected to compromise the DMS operation. The yields
expected in some parts of the orebody were higher than average, and if this
coincided with poor comminution and poor separation the throughput would be
severely limited by the capacity in the final recovery process. In addition to the
orebody grade, diamond assortment data had also been collected. These data were
used to fit a model to the revenue distribution within each size class that results
from the mix of diamond colour, clarity and shape within each size class. This
assortment model was used to quantify the influence of the range of the diamond
selling price on project cashflows.

The integrated evaluation model can test the impact of uncertainty related to the
orebody, the engineering design decisions (e.g. plant capacity) and future outcomes
(diamond selling price) and then combine the outputs in several different ways. To
clarify the use and operation of the model the terminology used is given in Table 36.

Stephen Coward - April 2020 203

 Evaluation of metallurgical recovery factors for diamonds recovered from kimberlites

1
Spatial
2
Realisations realisations of the
…
Orebody
25
1 Several Mine
Mine Plans
2 Plans
1
Alternative
2 Several versions
Iterations 3 of process
4 performance
5
1
Versions of future
Scenarios 2
outcomes
3
Table 36: Schematic depiction of terminology used to describe aspects of the Integrated
Evaluation Model.

Alternative - this term refers to the combination of inputs, processes and outputs
that are considered in the model. At this level it is possible to compare and contrast
very different approaches to operating strategy for a mining project.

Realisation - refers to one image of the orebody that honours the sampling data.
Typically, all realisations are processed through the model at a block-by-block scale.

Mine plans- there may be typically one to five mine plans that are developed based
on different assumptions including ground conditions, capacity, fleet size, angle of
repose etc.

Iterations - these refer to different operating conditions or efficiencies in the

mining and metallurgical processes.

Scenarios – Typically, three to five scenarios are developed to provide coherent

internally consistent futures that the project will possibly operate in. In each
scenario the trajectories of values that have a material impact on the project (e.g.
cost of capital, diamond selling price) are described, as are the relationships among
these inputs. These are most often financial inputs to the model but may also include
various environmental variables that have an impact on project performance e.g.
higher and lower rainfall impacting on mining rate.

Alternative - This represents a collection of plausible realisations, mine plans,

Iterations and scenarios. Each alternative would arise from a materially different
strategy of how the entire project is to be conceptualised and operated (e.g. large-
scale vs small scale, open pit vs combined underground and open pit)

204 Stephen Coward - April 2020

 Chapter 8: Value chain model - case study 2

The author was responsible for generating a number of orebody simulations,

incorporation of the mine plan, and development of VBA code to run the simulated
orebody models through the value chain. Grant Nicholas and Johan Ferreira
collaborated on this project which resulted in a conference publication (Nicholas,
2008). This project led to the development of a methodology to generate individual
diamond $/ct values by sampling from the modelled diamond value distribution.

Project background
The orebody is a kimberlite pipe located in Botswana. The tri-lobate pipe-like
orebody has intruded through the basement granites and is variably contaminated
by diluting basalt. This dilution is variable with maximum measured values of 45%
measured at 900m amsl in the north lobe and 25% measured in core retrieved from
the south lobe at 800m amsl.

The following sources of variability were considered to have important implications

in this project:

• The variability of the mined grade that would be achieved;

• Magnitude and type of internal dilution that will be encountered
when mining;
• The impact of hard waste and hard kimberlite on comminution circuit
and liberation performance;
• Impact of low comminution and dense minerals on DMS yield and
DMS separation efficiency;
• Capacity limitation in the final recovery; and
• Uncertain and variable diamond assortment impact on diamond
selling price

A number of mitigating steps were taken during the design phase to provide
sufficient flexibility in the process plant to cope with anticipated comminution and
yield challenges. Some of these design features included:
• Stockpile capacity ahead of the process plant, as well as a 150t surge bin
ahead of the recovery plant which would allow mining to change zones
should high yield be experienced;
• Optimisation of cyclone configuration to allow for higher spigot loadings
without compromising efficiency;
• Inclusion of capability to adapt the HPGR crushing circuit to cope with harder
material;
• Elimination of recycle of final concentrate by including grease belts and mills
in the final recovery flowsheet; and

Stephen Coward - April 2020 205

 Evaluation of metallurgical recovery factors for diamonds recovered from kimberlites

• Creation of flexibility in the layout of the final recovery to allow for inclusion
of additional units as technology and needs change - e.g. inclusion of rare
earth drum separators to deal with increased DMS yield.
Each of these risk mitigation features were designed to deal with the advent of one
of the risk factors and almost all the existing models were based on annual averages.
This meant that the existing project evaluation methods were unable to quantify
how well the project would respond to a combined synchronous 'onslaught' of all
the modelled uncertainty.

After consultation with the project team an integrated model was developed to
quantify the impact of various sources of uncertainty on project performance. The
next section describes the components of this model.

8.2 The value chain components

The capital for this project was limited and hence the project was based on a two-
phase operation, with the first phase of low throughput lasting approximately three
years, followed by a substantial expansion to increase throughput in phase two. This
would allow the project to acquire more information on the grade of the orebody
and to use the initial commissioning to gain greater insights into rock characteristics
and their impact on liberation and DMS separation. This approach to capital
rationing required a model that would be able to be adapted over time to reflect the
planned increase in mining rate and plant throughput.

The orebody models:

The surfaces of the orebody, consisting of three pipe-like bodies, were modelled
using conventional geological techniques (Figure 61). These surfaces were used to
constrain several simulations of the grade and density of the orebody. The turning
bands method of geostatistical simulation was used to generate 25 realisations of
the orebody.

The mining model

The mining sequence was designed to reduce capital expenditure and was split into
two phases. The first phase was planned to be a relatively small pit (down to 70m)
that would operate for the first three years and was expressed as pushback surfaces
per quarter.

The second phase targeted the southern lobe and was designed to extend down to
390m. This phase was based on annual surfaces that used tag blocks that were to be
mined in any period. A simplified sequence was derived based on reasonable

206 Stephen Coward - April 2020

 Chapter 8: Value chain model - case study 2

assumptions about directions and geometric constraints on the advance of the

excavation.

Figure 61: A view of the three lobes of the deposit looking from the west to the east (South
Lobe in Dark Blue) adapted after Campbell (2009).

The processing models

The phased approach to mining suggested that the initial plant design would be
required to treat 400tph. During years three and four the process plant capacity
would be expanded to 630tph and be adapted to improve the capability of handling
high yield kimberlitic material. Although the project team considered several
options during the feasibility study the process plant configurations remained
consistent Figure 62. The main flexibility in the design was the inclusion of the civil
infrastructure for additional streams required in phase 2, but the timing of
acquisition of the main capital items for these additional streams would depend on
the information acquired during phase one of the operation.

Stephen Coward - April 2020 207

 Evaluation of metallurgical recovery factors for diamonds recovered from kimberlites

Capital Item Number for phase 1 Number for Phase 2

Scrubbing and screening 2 4

Secondary Crushing 1 2

Fines DMS 1 3

Coarse DMS 1 2

Reconcentration DMS 1 2

Pneumatic drier 1 2

X-ray machines 4 8

Single Particle Sorters 0 1

Table 37: Summary of major capital items planned for each phase.

The design criteria for the sizing of these units was based on extensive studies of
core samples that were treated to determine the yield from dense media separation
that could be expected from various depths in the orebody (Table 38)

Depth Fines Module Coarse Module

0-70m 4% 2%

70-390m 37% 39%

Table 38: Average expected DMS yields for each phase.

Although these average yields were problematic the design team felt that it would
be uneconomic to design the plant in the early phase to cope with this high level of
yield. It was also important to consider that this yield was based on small samples
and the project required an understanding of the sequence of high yielding blocks
to evaluate the effect that stockpiling, and blending might have on ameliorating the
negative impacts of designing for high instantaneous yield.

208 Stephen Coward - April 2020

 Chapter 8: Value chain model - case study 2

Figure 62: Schematic of process flows used for process model.

The diamond assortment model

Given the relatively few numbers of stones that were used in the evaluation of the
expected diamond price that would be achieved, a model was developed to draw the
value from each stone from a calibrated distribution for value in each size class.

For every block in the mine plan, the grade and average stone size was used to
calculate the number of stones in each block. The mass of each stone was then
determined by drawing at random from a modelled cumulative stone size
distribution. In each mass class a cumulative distribution of the revenue in that size
class was used to draw a value at random for each stone. In this way the value for
the diamonds in each block could be cumulated to derive the realised value that
could be expressed as either the $/ct for diamonds in each block, or the $/tonne of
that block (Figure 63).

Stephen Coward - April 2020 209

 Evaluation of metallurgical recovery factors for diamonds recovered from kimberlites

Stone Size Simulation

8 300

7 Sim stone size cts

250
6 Average stone size
Average $/ct 200
5

Average $/ct
Size in Cts/stn

4 150

3
100
2
50
1

0 0
0 5000 10000 15000 20000

Stone Number

Figure 63: Plot showing a simulation of the size of 20 000 diamonds drawn.

The financial model

A conventional discounted financial model was built to value the project. The model
was set up in an excel spreadsheet and was run using VBA scripts. The model has
several variables that can be set prior to running the model. An example of the
settings used for the base case model is given in Table 39.

Model Asumptions Unit Value

Life of mine Years 14
North lobe Diamond Price Growth Annual Factor 1.03
Central lobe Diamond Price Growth Annual Factor 1.03
South lobe Diamond Price Growth Annual Factor 1.03

Annual
Costs Unit Base Escalation
factor
Ore Mining cost US$/tonne 3.02 1.1
Waste Mining Cost US$/tonne Waste 2.97 1.1
Ore Processing cost US$/Tonne 8.62 1
Ore Treatment Cost US$/tonne 27.32 1
Carat Recovery Cost US$/ct 188.69 1

Royalties Unit Value

Administration Overheads Percent of Revenue 2.5

Process Recovery Modifier Unit Value

North lobe Unitless 1
Central Lobe Unitless 1
South Lobe Unitless 1

Table 39: List of model settings used for the financial model.

210 Stephen Coward - April 2020

 Chapter 8: Value chain model - case study 2

These settings can be varied individually or as a group. The scenario approach to

evaluation provides a basis for qualitative correlation between the settings, i.e. if the
selling price is high and growing it is likely that the input costs would also be rising.
This approach can be used to limit the number of times that the simulation model
must be run.

8.3 Considerations for development and simulation of integrated

value chain models

The development and implementation of a dynamic and robust value chain

simulation requires a flexible and rapidly configurable modelling environment.
There is also a requirement to provide users with the ability to re-run the model
with a variety of adapted settings and save the associated results in a structured and
sequential framework. This system was thus designed using the concepts that
underpin the model view controller architecture (MVC) as depicted in Figure 64.

Integrated Value Chain Model

Action
Instruction Manipulates
Controller

User Database

View Model
Update Notify

Figure 64: Schematic of the Integrated Evaluation Model Architecture (adapted after
Burbeck, 1992) .

In this architecture the interaction of the user with the application is separated from
the control of the model and access to the data by a so-called "View" object. The
"controller" object interprets user instructions that are sent from the view to the
controller. These instructions are then routed to the "Model" object. This model
object is the heart of the simulation and is capable of sending multiple structured
requests to the database, as-well as sending outputs to the database. When the
model has completed an instruction, it notifies the "controller" object. Depending
on the logic embedded in the controller it may then either re-manipulate the
"model" or send an update to the view or do both.

There are several benefits of this approach to development including:

Stephen Coward - April 2020 211

 Evaluation of metallurgical recovery factors for diamonds recovered from kimberlites

• This view object can be readily changed and adapted as the application
evolves and acquires additional functionality. It is also possible to obscure
functionality from the user until it is ready to be included;
• Ability to separate the development of the individual components, so that
each component can be developed and tested separately. At the point where
the model is to be integrated, the logic that controls the interaction between
the components can be designed, tested and implemented. This sequence
reduces the complexity of detecting and eliminating errors from the system;
and
• Development of parallel processing architecture, in the IEM case each
combination of orebody and value chain can be separated and run
simultaneously by the "model object" which is far quicker than running each
of these sequentially.
The "view" was developed in Microsoft Excel as it provided a familiar and readily
customisable user interface. The "controller" was written in Visual Basic for
Applications in Excel. Several commercial applications were used to support the
sub-models that underpin the IEM framework e.g. Isatis for spatial simulation,
Statistica for generation of stone distributions, Datamine for mine plan sequencing.
Microsoft SQl express was used as a database for storing input data, processing
parameters and result data.

The orebody models can be quite large and exist in a multitude of formats in the
software that is used to generate them. They were thus exported from their source
software and stored as flat text files. These files were in turn read into the database
with a table for each realisation. Likewise, the mine plans were contained in
separate text files, with each block having been assigned an ideal year for extraction,
which was brought into the database. Lists of blocks to be processed in each year
were generated by executing a dynamic query that was triggered by the process
model. The process model used the simulated block properties to determine the
treatment time and the liberation that would be achieved on a block-by-block basis.
In each year the treatment ceased when the allocated hours available in a year had
been consumed. The production outputs were written back to the database and the
next year's list of available blocks were retrieved. The production consumption data
and diamonds produced were carried forward to the financial model to determine
the costs and revenue in each year.

The use of an integrated system meant that post initial configuration and testing, a
standard i7 processor desktop computer was able to process a set of five
alternatives (each alternative consisting of 25 realisations, four iterations and three
scenarios) in just over eight hours.

212 Stephen Coward - April 2020

 Chapter 8: Value chain model - case study 2

8.4 Case study findings

In this case study the block models contained more than just grade information.
Each block also contained variables that could be used to determine the
comminution, liberation and DMS efficiency and the results of spatial simulations of
diamond size distribution and value within each size class. The project also required
differing treatment rates to be achieved over the life of the operation.

The aim of this case study was specifically to identify processing challenges and how
these could be addressed through pre-emptive risk mitigation that would not
compromise the capital constraints placed on the project.

The summary statistics of the simulated variables in the mined blocks are presented
in Table 40.

Grade Density Yield

Count 62975 62975 62975
Minimum 0.00 1.80 0.06
Average 0.51 2.59 4.63
Maximum 6.42 3.17 48.00
Stdev 0.342 0.311 5.830
Coeff Var % 67.6 12.0 125.9
Table 40: Summary statistics of spatially simulated Grade, Density and DMS yield
values.

The yield value has the highest coefficient of variation followed by the grade and
then the density. This suggests that the uncertainty in the yield is higher than that
for grade. The apparent low range in density values is partly a result of inserting a
background value of 2.90 for blocks to be estimated that were further away from
samples than the range of the variogram.

Throughput
The throughput model planned for an annual mined capacity of 600kt per quarter
and a process plant or mil capacity of 525 Kt per quarter. A summary of the tonnage
treated is given in .

In the initial period the ramp up for mining and plant increases markedly from year
one to year two. The plant capacity is satisfied from the middle of the project
onwards. During the ramp-up period however the cash flows derived from
production fall short of the plan. This is primarily an impact of the slower rates
achieved due to the processing of harder ore.

Stephen Coward - April 2020 213

 Evaluation of metallurgical recovery factors for diamonds recovered from kimberlites

Figure 65: Summary of tonnage treated

In the initial period the ramp up for mining and plant increases markedly from year
one to year two. The plant capacity is satisfied from the middle of the project
onwards. During the ramp-up period however the cash flows derived from
production fall short of the plan. This is primarily an impact of the slower rates
achieved due to the processing of harder ore.

Revenue from diamonds recovered

The simulation of stone sizes and values generated a total of just over 600 M stones
for the life of the mine and took approximately four hours to generate these stones.
The stones were seeded back into the orebody to produce a range of $/tonne values
for each block as it was mined. The summary statistics for these data are given in
Table 41.

214 Stephen Coward - April 2020

 Chapter 8: Value chain model - case study 2

Lobe Minimum Average Maximim Stdev Coeff. Var. %

N 11.32 36.39 87.83 16.00 43.96
C 11.22 32.24 52.84 8.42 26.13
S 2.51 18.70 53.01 6.11 32.66

Table 41: Summary Statistics of $/tonne depleted by lobe based on simulated stone
values.

The highest $/tonne value, on a mining block scale, was 87.83 $/tonne in the North
lobe which also showed the highest coefficient of variation. This was to a large
degree driven by the coarser stone size distribution in this lobe and the contribution
of larger stones in the assortment.

Project Valuation

The initial modelling identified that using a 10% discounted cashflow model the Net
Present Value(NPV) ranged from BWP -262 million (P10) to BWP +1 million(P90)
with a P50 NPV of BWP -156 ( Figure 66) .

The model identified that the cashflows in the plan for 2010 and 2012 had the
highest uncertainty, and that the project team needed to review mining and process
flexibility in these periods. A shortfall of the order of 10% of tonnes treated in this
period of the project would translate to close to a 10% change in project value.

Stephen Coward - April 2020 215

 Evaluation of metallurgical recovery factors for diamonds recovered from kimberlites

8.5 Generic findings from the IEM relevant to the derivation of

the metallurgical recovery factor

Figure 66: Summary Plot of Cumulative Discounted Cashflow for the Ak project, P50 case
shown in green,P80 and P20 case shown in red, individual cases shown in grey.

The traditional approaches to deriving project value using single smoothed orebody
models are likely to under-estimate the impact of orebody variability on project
value. The integrated modelling approach facilitates a far richer exploration of the
impact that orebody uncertainty will have on project performance. The same
potentially biased estimation of recovery factors can also result from the use of
single averaged assumptions about process performance.

It is possible to arrive at an expected recoverable diamond distribution using a

truncated average total diamond content curve. The truncation is achieved by
applying a fractional factor to stones that are expected to be recovered below the
lower cut of screen size. This modified diamond stone size distribution can then be
used to represent the expected recovered diamond distribution and concentration
for the life of the operation. Figure 67 shows how it is possible to truncate the total
content curve to represent the impact of changes in bottom cut-off size used in the
process plant on the expected size distribution of recovered distributions.

216 Stephen Coward - April 2020

 Chapter 8: Value chain model - case study 2

Figure 67: Grade size plot showing the impact of applying a strict size cut-off to a total
content curve.

In this particular case a combination of bulk sample and microdiamond results were
used. This approach applied to this project gives a range of total content carat
recovery ranging from 0.84 at 1mm down to 0.51 at 2mm bottom cut-off size. Using
the average grade of the resource and its average distribution this would translate
to a revenue recovery range of just over 3 $/tonne. This clearly underestimates the
variability that is likely to be experienced during the life of the operation as shown
using the integrated model.

Using this approach it is possible to define the recovery factor in several alternative
forms, working back from the bulk sample recovery factor (Equation 75)to the
intended main treatment plant recovery factor(Equation 76) . The ratio of the
intended plant recovery factor to the bulk sample plant recovery factor is then used
to determine the resource to reserve factor that is used to account for metallurgical
efficiency (Equation 77).

𝑩𝑺𝑷𝑹𝒆𝒄 𝑭𝒂𝒄 = 𝑩𝑺𝑷𝑳𝒊𝒃 𝑭𝒂𝒄 × 𝑩𝑺𝑷𝑭𝒓𝒆𝒆 𝑫𝒊𝒂𝒎𝒐𝒏𝒅 𝑹𝒆𝒄 Equation 75

𝑴𝑻𝑷𝑹𝒆𝒄 𝑭𝒂𝒄 = 𝑴𝑻𝑷𝑳𝒊𝒃 𝑭𝒂𝒄 × 𝑴𝑻𝑷𝑭𝒓𝒆𝒆 𝑫𝒊𝒂𝒎𝒐𝒏𝒅 𝑹𝒆𝒄 Equation 76

Stephen Coward - April 2020 217

 Evaluation of metallurgical recovery factors for diamonds recovered from kimberlites

𝑴𝒆𝒕 𝑹𝒆𝒄 𝑭𝒂𝒄𝑹𝒔𝒄𝒕𝒐𝑹𝒔𝒗𝒆 = 𝑴𝑻𝑷𝑹𝒆𝒄 𝒇𝒂𝒄 ⁄𝑩𝑺𝑷𝑹𝒆𝒄 𝑭𝒂𝒄 Equation 77

Using this approach with average bulk sample plant recovery as the input, the range
of the metallurgical (resource to reserve) recovery factor was calculated to range
from 0.97 to 1.08. When using the integrated evaluation model approach, it became
apparent that the variability due to processing constraints was far higher. The
variability produced by the IEM correctly reflects the combined impact of the
variable grade of the ore being treated and the variable efficiency of the process that
results from the changing characteristics of the rock treated. This approach also
allows a much finer temporal scale evaluation of variability in throughput, total
carat production and hence project value. Using the IEM approach, it is also possible
to assign a localised ore recovery factor and in doing so identify where in the
orebody the recovery is likely to vary most, and why it will vary.

The rock characteristics developed here were mainly content variables, however
their impact on liberation of dense species within the ore also impacts on
throughput and recovery quality.

The use of a log normal stone characteristic simulation allowed the model to
incorporate the uncertainty of diamond assortment on the expected recovered
revenue. Even with a small number of parcels of stones it is possible to demonstrate
the range of expected revenue that extends several orders of magnitude beyond that
achieved with a simple static cut-off analysis.

This case study has demonstrated that it is important to consider variable small-
scale (mining block scale) impacts of rock properties, rather than the average
impact of average kimberlite characteristics, to determine correctly the range of
metallurgical recovery that will be experienced over the life of the operation. This
requires multiple images of the orebody to be processed at a block scale through a
dynamic integrated value chain model. This method is feasible and relatively
straight forward to implement.

The final chapter concludes with a summary of the developments that this research
has provided in the derivation and use of metallurgical recovery factors for project
evaluation.

218 Stephen Coward - April 2020

 Chapter 9: Discussion

9 DISCUSSION

9.1 Introduction

This thesis describes current practices used to derive and evaluate metallurgical
recovery factors for kimberlitic diamond mines. The recovery factors are used to
account for the discrepancy between in situ values of variables that drive project
value (diamond stone concentration (stones/ht), diamond grade (Ct/ht), diamond
size (Cts/stone), diamond value($/ct), and ore value ($/tonne)), and the expected
'recovered value' of these variables. The literature reviewed shows that predictions
of recovery that are based on assumptions of average rock characteristics acquired
from a few spatially dispersed samples can be biased.

Several significant improvements to the current approaches have been presented in

this thesis and they have been demonstrated to give better recovery factors. These
include:

• Clarifying the variable taxonomy using the primary response framework;

• A methodology to develop proxy measurements to augment costly time-
consuming destructive rock characterisation testing;
• The use of multiple data sources to generate combinations of estimated and
simulated orebody models of rock characteristics;
• The development and use an integrated evaluation framework to model and
simulate process performance;
• The application of uncertainty and range analysis techniques to demonstrate
how the range of expected recovery factors can be used to mitigate project
risks.

The application and benefits of each of these improvements are briefly described
below.

9.2 A taxonomy for variables used in recovery factor

determination

The building of spatial models of orebody characteristics that can be used to derive
recovery factors requires data with sufficient spatial intensity and appropriate
support to characterise their spatial behaviour. Historically, sampling for rock

Stephen Coward - April 2020 219

 Evaluation of metallurgical recovery factors for diamonds recovered from kimberlites

characteristic variables that can be used to predict recovery has often been limited
to a few large 'representative' samples that do not capture the range or the spatial
variability of the characteristics of interest.

To overcome some of the physical and financial constraints of sampling, several

innovative technologies and approaches have emerged. These include the
augmentation of high-cost large support samples with smaller, less costly samples
or measurements that can be taken from more places in the orebody. These so-
called 'proxy' tests or measurements are typically slightly less accurate than the full-
scale representative sample and require the building of some form of mathematical
model to enable the proxy data to be used in the generation of the spatial models.

There is a requirement for a taxonomy to classify the numerous tests, data sources
used to quantify the variables used in recovery factor evaluation. An ideal taxonomic
framework will guide the selection of samples and the tests that they are subjected
to so that the data sets generated can be effectively used. The suggested taxonomy
provides a useful framework for design and implementation of the approached
described here. Variables used in recovery factor evaluation range from diamond
content variables to characteristics that describe the primary properties and
responses of treated rocks to some energy input (Figure 68). The use of two such
approaches (Coward et al, 2003) and (Keeney and Walters, 2008) are described to
demonstrate how they can be used to balance the rationalisation of sampling with
the need for sufficient data of appropriate support to facilitate spatial estimation
and simulation of the characteristics of interest.

Figure 68: A schematic representing the primary response framework for geometallurgical
variables.

220 Stephen Coward - April 2020

 Chapter 9: Discussion

The primary response framework for geometallurgical variables provides a useful

distinction between variables that are closely associated with the in situ rock
characteristics and variables that are associated with the response of the rock to
some process. Primary variables typically are easier to estimate but require more
spatially intense sampling. Response properties are often characterised by
nonlinear relationships which make them harder to estimate and require larger
samples with larger support. The third dimension in this classification - 'energy and
process' - gives an indication of the meta data that should be collected when carrying
out measurements and tests on rock samples.

A taxonomy developed by Keeny and Walters (2008) suggests there are four levels
of data (Figure 69), with level one being geologically focussed and spatially
representative and level four data being typically acquired from large composite
samples that have a metallurgical focus. This suggests that Level 1, 2 and 3 data are
derived from drill core, and in some cases from large diameter drilling (LDD) that
produces rock chips. On operating mines it is possible to gather additional level 1-3
data from in situ samples taken from the pit, or from post production blasting and
during mineral processing.

Relationship to Spatial Coverage

Processing
Traditional Metallurgy
Cost/Sample Very Limited
Large-scale composites Level-4
Direct
Discontinuous narrow
Geometallurgy

intervals of core for

Level-3 Limited
physical testing
Direct Proxies
Continuous (?) core-
based new Level-2 Good
measurements Indirect Proxies

Existing core-based Level-1 Very Good

No. Samples
continuous databases
Traditional Geology

Figure 69: A data typology adapted after Keeney and Walters (2008).

In the case of sampling to measure the characteristics of kimberlites and their

diamond content, their particulate nature and the low concentration of the
mineralisation generates some data types that do not quite fit into this taxonomy.
For instance, large samples of several thousand tonnes are occasionally collected
and treated at operating mines (and in some cases from trenches made across
orebodies in early project phases) to obtain a sample large enough to represent the
contained diamond footprint (ranges of colours, shapes and qualities). In this case
the sample has very limited spatial coverage, is discontinuous, and the main

Stephen Coward - April 2020 221

 Evaluation of metallurgical recovery factors for diamonds recovered from kimberlites

objective of collecting the data is focussed on a content variable that is arguably

geological in nature.

Although the classification has some limitations in kimberlites, it is, however,

relatively simple and provides a useful framework for developing a sampling
programme that aims to achieve a balance between the high-cost per sample and
low coverage of level 4 data, with the lower cost per sample and higher coverage
that is associated with collection of level 1 data. This type of trade-off is often
confronted by mining project evaluation programs. The 'optimal' sampling
programme depends on a number of factors that are specific to the characteristics
of the kimberlite to be mined and the mining and process configuration. The best, or
optimal, combinations of number of samples, sample support and tests carried out
on each sample will differ from mine to mine, deposit to deposit and lithology to
lithology and will be specific to the variables required to inform the project as-well
as the phase that the project is in.

By ranking and mapping the planned tests and sampling on an independent

classification for sample methods based on their level (Keeney et al., 2008) of
coverage and the degree to which the sample and tests are used to derive data that
reflect a primary or response property of the variables it is possible to evaluate
which trade-off will make most sense in terms of time, cost and the ability to
generate models and reconcile the model to make useful value adding decisions in
both the tactical and strategic aspects of the mining business.

Figure 70: A landscape for sample type classification in terms of both spatial continuity and
primary response dimensions.

In Figure 70, a landscape for classification of tests is presented - the scales on the X
and Y axes are broken down into two regions - giving four quadrants into which
each sample type can be placed.

222 Stephen Coward - April 2020

 Chapter 9: Discussion

These taxonomies suggest that there is substantial benefit in applying several

generic principles when designing and implementing a sampling programme that
aims to acquire data that will support the estimation of metallurgical recovery
factors. Table 42 Presents a potential list of such guidelines that will be very useful
in defining the parameters for designing sampling experiments and subsequent
workflows for the purpose of providing sufficient information to facilitate the
evaluation of a project or mine metallurgical recovery factors.

Principles Explanation
Spatial coverage The data used for geometallurgical modelling should be spatially distributed
and on an appropriate support to permit reasonable estimation of the chosen
variables, including domaining and selection of estimation techniques.
Primary Response Rock characteristic variables can be usefully classified as either primary
Variable Classification attributes that are closely associated with in situ rock characteristics and
response variables that are closely related to the response of the process to
rock characteristics. The relationship between primary and response variables
is also a function of the process used to generate the response.
Sample rationalisation Sample selection and analysis are rationalised to meet the demands of being
both adequately representative of geology and processing responses. For
example, fewer samples will be required for early exploration of a kimberlite
to inform the decision to continue sampling, than the number of samples
required to develop a model that can be used for resource classification.
Co-location As far as practically possible, samples chosen for analysis should be co-
located in space and at a scale that is appropriate and relevant to the model
that is being built and the intended use of that model. Isotopic data, i.e. having
as many of the same measures on samples at all sampled locations, have value
in their potential to be used for development of proxy measures.
Calibration and new Where response variable proxies are being proposed to predict primary rock
technology characteristic variables, calibration of the relationship between the proxy
variable to primary characteristics must be undertaken at an appropriate scale.
This requirement is applicable to both existing and new technology.
Business case An appropriate business case should be made for any sampling campaign
with respect to the tactical and/or strategic model that is to be designed and
created and used to support decision making.
Generic vs specific Generic guidelines are considered useful for sampling problem framing and
sample to model planning and development at a strategic level. It is important
that specific site, domain and variable specific sampling experiments should
be designed given the heterogeneity of kimberlitic deposits.
Objectives and The objectives of each sampling campaign should be clearly articulated, and
measurements measurement criteria defined to assess the success or failure of each
campaign
Principles of correct Correct sampling practices should lead to unbiased results. Design and
sampling execution of any sampling experiment must consider good sampling science
(Gy, 2004) and be designed to minimise sampling bias.
Building on existing Where possible, new sampling campaigns must allow existing data to remain
data valid so that the database grows organically, and that no repeat sampling is
required; e.g., consistent use of a set grind size for slime content analysis.
Table 42: A list of guiding principles for rock characteristic sampling.

Stephen Coward - April 2020 223

 Evaluation of metallurgical recovery factors for diamonds recovered from kimberlites

These guiding principles provide a useful framework to plan and execute a sampling
programme that will produce a data set of sufficient quality for the purposes of
spatially estimating rock characteristics that can be used to evaluate metallurgical
recovery factors.

9.3 Increasing the quality of spatial estimates -using proxy

measurements, scale-up and testing for additivity or non-
linearity

Prior work on increasing spatial coverage of rock characteristic data (Dowd, 1997)
has shown that it is possible to infer rock characteristics at unsampled locations
using a combination of direct and indirect data. This research has extended this
concept and presented an experiment to demonstrate how a similar approach can
be used in Kimberlitic rocks. The experiment has shown how it is possible to acquire
data sets from a combination of in-hole measurements, measurements on core and
regional geophysical measurements. These data make it possible to improve the
qualitative understanding of geological domain boundaries and create a framework
for finding proxy measurements for difficult and costly destructive rock tests.

The collection of the data requires attention to be paid to descriptions of the core
and careful colocation of the measurements made down the hole and on the core. It
was found in this specific data set (i.e. a southern African volcaniclastic kimberlite
and kimberlite breccia) that averaging to approximately 70cm increased
correlations between various measurements. This result suggests that the scale of
geophysical response and errors in location are minimised at this level of support -
for the rock types tested here. In applications to other deposits a similar approach
will be required to improve the correlations between the response measures made.
The methodology of finding the minimum point of the summed within-sample
variance and between sample variance can be applied to several variable pairs and
could produce different ideal support sizes.

Work described here has identified that scale-up is not complex where variables
exhibit linear relationships and can be considered to have additive behaviour. The
approach required for variables that exhibit non-linear relationships, of which there
are many in this field (e.g., bond work index, abrasion index), are likely to require
more considered, and potentially far more complex approaches (Carrasco, 2014). A
pathway for developing suitable approaches for spatial modelling of a variety of
rock characteristic variables is discussed in chapter 6. This approach suggests that
the parameters used to generate the spatial model are not independent of the use
for which the model is intended. For example, in some cases the when generating a
kriged grade estimate, the modeller may choose a set of kriging parameters that will
be better suited to providing an estimate that is more suited to a global or local

224 Stephen Coward - April 2020

 Chapter 9: Discussion

estimate. The methods used to generate spatial simulated models of rock

characteristics can also be tailored to the intended use of the spatial simulations (for
example by choosing appropriate block sizes and setting up of domains that reflect
the spatial nature of the physical rock characteristic being simulated). Block to
block variability of the rock characteristics may be important when the spatial
models are used in highly granular dynamic simulations to predict rates and
efficiencies of mining and metallurgical recovery processes.

9.4 Diamond metallurgical recovery factors – exploring

uncertainty and variability by effective use of value chain
models

The linking of populated multivariate models of the orebody with process models
that respond at an appropriate scale opens several avenues to derive and evaluate
the Metallurgical Recovery Factor for Diamonds recovered from Kimberlitic
Deposits.

In early stage projects where there is little information on the orebody, and still
significant flexibility in design criteria selection, it is still useful to have a
quantitative model that links the orebody characteristics to the expected recovered
diamond population. Although the expected value for total recovered diamond
mass, size distribution and revenue distribution are important, it is perhaps more
important to be able to develop and apply a methodology to quantify the ranges of
these values.

The published literature has demonstrated that it is common practice to use the
average rock properties either at a global scale or at a domain scale to predict the
fracture and process rates of rocks through a given process (Farrow, 2019). This
limited approach is, in some cases, justified by the limited availability of rock
characteristic data and the inability of traditional approaches to incorporate this
information effectively in the derivation of recovery factors.

System theory (Boulding, 1956) suggests that dynamic effects that arise from the
interaction of constraints with variability will not be correctly reflected in models
that are based on long-term averages. This same effect is referred to in the theory of
constraints (Goldratt, 1990). In this theory, the average of the throughput of several
interdependent processes will not equal the average throughput. This also known
as the 'flaw of averages' (Savage, 2008) and although it seems that these biasing
effects would be explicitly addressed in mining project evaluations, it is evident that
several projects are evaluated on data that are summarised on an annual scale for
inclusion in financial models. The biases that arise from this shortcoming are
difficult to detect and may be very complex. It is for these reasons that a highly

Stephen Coward - April 2020 225

 Evaluation of metallurgical recovery factors for diamonds recovered from kimberlites

granular interdependent system model of the mining and treatment process should
be used to evaluate the expected range and variability of performance of the
diamond recovery processes.

9.5 Conclusion

This research aimed to investigate ways to identify and measure the rock
characteristics that have an impact on the metallurgical recovery factor. A case
study has shown it is possible to sample for several response characteristics and
then estimate these effectively using a so-called proxy framework to augment
response property data with non-destructive measurements in the hole and on core.
These data also facilitate the generation of several spatial simulations of these
characteristics. The development of a way to process these multivariate models of
the orebody at a fine scale (block scale resolution) enables the assessment of the
impact of these rock characteristics on diamond recovery on widely varying time
scales. The output data can be analysed in many varied ways to quantify the mean
and the range of recovery factors for diamonds recovered from kimberlitic deposits.

Data processing speed has increased exponentially for several decades (Moore,
1965). Effective access to this increasing rate of processing has to some extent been
limited. Recent developments in arbitraging rates for excess server capacity by
some of the broadcast networks, via various service providers, including Amazon
Web Services and Oracle, suggests that the constraint on processing large and
previously unwieldly data sets is becoming affordable to mid-tier mining
companies. This suggests that the cost of providing platforms for processing large
multivariate orebody models will reduce and their use in the mining industry will
become more common. These developments mean that the methodology presented
here is not only valid but can now be practically implemented at a cost that is of a
similar scale to those traditionally associated with resource estimation and project
evaluation.

A large non-technical challenge to the adoption of this methodology has been the
ability to convey the outputs of the range analysis in a way that is embraced by
senior stakeholders in the mining industry. Although the concept of issuing
"Guidance Ranges" for critical performance targets is being more widely adopted
(e.g. Newmont annual report, 2015) it's use in corporate decision-making requires
more work but is gaining acceptability. Cyclicity, decreasing margins, lower average
grades and increasing competition for fewer resources are hall marks of the mining
industry. The development of methods that can validly assess, and respond to, the
impacts of uncertainty and variability of project value will confer special
competitive advantages on the companies that embrace it. The methodologies
presented here, when used as part of a strategy for improved decision-making, will
add substantial value to the mining industry.

226 Stephen Coward - April 2020

 Chapter 10: Summary and Conclusions

10 SUMMARY AND
CONCLUSIONS

10.1 Introduction

The specific limitations of traditional recovery factor evaluation that have been
addressed in this thesis are those associated with the difficulty of incorporating
variable and uncertain rock properties in the derivation of the metallurgical
recovery factors.

Metallurgical recovery factors are required to account for the difference between
the estimated in situ diamond content and expected diamond recovery. These
factors are used to modify the stone concentration, the mass concentration and
diamond size frequency so that the expected recovered $/tonne can be predicted.
Failure to explicitly include the impacts of rock properties in traditional project
valuation can result in sub-optimal mine design, incorrect processing configuration
(design, operating strategy) and potentially biased production and cashflow
forecasts. Not only are there vast differences between the scale of measurement and
the scale of estimation of these properties but the data acquired from tests and
measurements on small samples may not directly correlate with the rock-process
relationship that will eventuate at full-scale operation. The consequences of
potential biases are amplified when global estimates of rock and diamond
characteristics, based on few and spatially sparse data, are used for planning and
designing the entire life of the project.

The problem statement is described as:

• Current known estimation techniques for recovery factors of kimberlite

diamond deposits are based on global assumptions and average values and
do not explicitly incorporate the spatial variation of rock properties. The
problem stems from the following inadequacies:
– No existing method to measure and estimate the physical spatial
characteristics of kimberlite.
– Mine production and metallurgical process modelling is insensitive
to spatial estimates of physical character.

Stephen Coward - April 2020 227

 Evaluation of metallurgical recovery factors for diamonds recovered from kimberlites

– Project decision making/valuation frameworks do not incorporate

the change in project value that can be attributed to responses to the
recovery factor uncertainty.

A research framework was developed to address the problem statement that

focussed on three areas of research. The first area required a combination of in-field
sampling experimentation, and considerable work to develop a model to generate
multiple spatial realisations of orebody characteristics. The approach explored
methods to use multiple sources of data to create rock characteristic estimates.

The second area of successful research identified aspects of current practice in

process plant simulation that could be used to process the multiple realisations of
the orebody and predict process outcomes. Several methods were developed and
demonstrated in two case studies; one for an underground mine and a second for
an open-pit operation.

The third area of research was focused on a means to evaluate the range and
uncertainty of the derived forecasts for metallurgical recovery. This required the
development of an integrated value chain model that could be used to translate the
technical uncertainty into financial metrics. This approach allows for comparisons
to be made between various project risk mitigation strategies.

The remainder of this chapter highlights areas of successful developments, reviews

some of the limitations of this research and concludes with a brief discussion of
areas for further research.

10.2 Concluding Themes

This research has demonstrated that several current reviewed methods of

evaluating the metallurgical recovery factor for diamonds recovered from
kimberlite deposits have limitations. Improvements in several areas can now be
made as a direct result of the work reported here. These are briefly discussed below.

Variable selection and identification

The primary response framework (Coward et al., 2009) was developed to clarify the
taxonomy for variables that are used in this research. This conceptual framework
provides a basis for developing quantitative models of the relationships between
variables that drive uncertainty in the recovery of diamonds. The benefit of this
framework is that it provides guidance in the use, and limits to some extent potential
misuses, of methods for generating spatial models of rock characteristics.

228 Stephen Coward - April 2020

 Chapter 10: Summary and Conclusions

A study of several existing plant processes has identified important metallurgical

response areas and developed relationships between rock properties and the
performance of those processes. These can broadly be divided into characteristics
of the rock that impact on comminution, dense media separation, final diamond
recovery, water management and materials handling. The study also identified
several diamond properties that impact on their likelihood of being recovered.
These include diamond geometries, surface characteristics and luminescent
properties.

These insights have facilitated the development of an approach to improving the

design of sampling campaigns so that they provide enough data of sufficient quality
to facilitate rock characteristic estimation.

Sampling for rock characteristics

As a result of historical constraints placed on the acquisition and treatment of

samples to provide data to model unit process performance, the models have in
several cases evolved to consider so-called ‘average rock characteristics’ as their
inputs. This has in turn led to the use of so-called composite representative samples
that aim to emulate the average characteristics that will be encountered in the
treatment process. This approach has been shown to have several limitations that
can be overcome by developing sampling campaigns that acquire samples from
several locations with sufficient support to facilitate spatial estimation of the
characteristics of interest at unsampled locations. Where this can be achieved, the
selection and definition of geometrically constrained domains, so-called
“processing” or “geometallurgical” domains, becomes an important step in being
able to estimate these characteristics with any confidence. The requirement for
spatial stationarity becomes even more onerous when spatially simulating these
characteristics.

The benefits of this approach to establishing the spatial nature of rock

characteristics is that the estimated models will provide an unbiased estimate of the
expected characteristic at all estimated locations, and the spatial simulation
approach provides a pathway to explore and define the range of rock characteristics
that will be encountered when mining the deposit. As a suite, the simulations
provide a probabilistic expression of the range of outcomes that could be
encountered.

Sampling for rock characteristics requires a rigorous approach to sampling to

eliminate a host of biases that can arise – an approach has been demonstrated here
using an experiment carried out on an operating kimberlite mine.

Stephen Coward - April 2020 229

 Evaluation of metallurgical recovery factors for diamonds recovered from kimberlites

Kimberlites comprise a complex suite of rock types that exhibit a wide range of
physical characteristics. These arise from a combination of differences in
composition, texture and weathering or alteration state. Accounting for these three
aspects of the rock require suitable adjustments in the acquisition of samples and
the testing of characteristics of those samples. This is especially true for coarse
textured breccias that require a larger sample support. To mitigate some of the risks
associated with the constraints on the number and spatial coverage of physical
samples it is possible to use more cost effective, and potentially less accurate, proxy
measurements to enhance the estimation of values at unsampled locations (e.g.,
using acoustic velocity to augment various measures of rock strength that are used
to predict fracture.)

Spatial estimation and simulation

Spatial estimation and simulation are different methods that both produce spatial
models but that have differing characteristics that are useful in certain
circumstances. Both approaches require attention to spatial coverage, support and
scale-up, and provide different benefits to the evaluation process.

When sufficient data have been acquired it is possible to derive parameters for the
spatial estimation and spatial simulation of the rock characteristics of interest. Best
linear unbiased estimation methods such as ordinary kriging provide models that
are unbiased but potentially smoother than the reality that will be encountered
when mining. Conditional spatial simulation methods provide a means to replicate
the in situ variability of the characteristics of interest. As a suite, the realisations can
be used to determine the uncertainty of these characteristics. This is important and
there are often constraints in the mining and processing of these rocks that respond
to extreme values rather than the average characteristics.

Mining and treatment process simulation

Process simulation can be used in combination with spatial rock characteristic

models to quantify the impact of rock variability on diamond liberation and
recovery. A model to do this has been developed and implemented in this research.

Conventional unit process models have traditionally been limited by both speed of
processing and the limited availability of valid spatial models of characteristics that
could be used as inputs for dynamic process simulation. The evolution of a
methodology for using spatially estimated (and spatially simulated) rock
characteristics to simulate total comminution and link total comminution to
diamond liberation and diamond recovery provides a quantitative way to forecast

230 Stephen Coward - April 2020

 Chapter 10: Summary and Conclusions

the mean and expected range of metallurgical recovery for a given process
flowsheet.

A linked value chain model is an appropriate way of evaluating the impact of rock
and diamond variability on the expected value and range of the metallurgical
recovery factor. The benefits of an integrated value chain model include:

1 A quantified approach to estimating the range and uncertainty in the

metallurgical recovery factor will improve the identification and
selection of strategies to optimise kimberlite mine operation.
2 As the model uses multiple spatial realisations of the orebody
characteristics it is possible to identify periods when the recovery
factor is most uncertain and hence identify which actions (e.g.
additional sampling, or additional crushing capacity) will have the
biggest impact on metallurgical recovery.
3 The process model is calibrated to a specific flowsheet and can thus
be adapted and rerun to determine how changes in design and
operation will impact on the metallurgical recovery. This allows for
iterative optimisation that targets robust recovery rather than a
vulnerable ‘optimum’ design that does not explicitly consider
uncertainty.

10.3 Specific contributions of this research

This research has specifically addressed the following areas:

• Development and demonstration of an approach to extend the spatial

coverage of data that is acquired from costly destructive rock testing
methods;
• Investigation and demonstration of viable methods to estimate and simulate
rock properties into block models at an appropriate scale;
• Development of an integrated value chain model that facilitates analysis of
the impact that variable and uncertain rock properties have on mining and
treatment efficiency; and
• Use of an integrated evaluation model to calculate the metallurgical recovery
factor for the life of diamond mining projects at different resolutions, and to
generate confidence limits around the estimate of the metallurgical recovery
factor.

Stephen Coward - April 2020 231

 Evaluation of metallurgical recovery factors for diamonds recovered from kimberlites

10.4 The limits of this research

Uncertainty and variability are two very distinct characteristics of a variable.

Unfortunately, the variability of any variable, as described by its values, is a
composite of several potential sources of variability. The spatial models and
simulations of the values of these variables at unsampled locations will produce
realisations with variability that is a function of both the true spatial variability and
the variability that is introduced by other sources of variability from the measuring
and testing processes. Understanding the relative magnitude of the variability that
is intrinsic to the deposit and the variability that arises from the uncertainty due to
the inability to measure accurately at all locations will require further work.

The fitting of functions to data imposes the modellers will, not only on the
parameters that are used to control the location and shape of the function used, but
in the selection of the form of the equation that is used to resemble a real-world
phenomenon. Some of the functions that are used in mineral processing models
have been validated by comparing their predictions, on average, with the averages
of observations of real-world process plant responses (e.g., exponential
comminution models). The approach developed and demonstrated in this thesis
uses errors, or residuals, of model fitting to generate a feasible range of the
parameters for the models used. One such approach used here allows the inclusion
of the standard error to induce a measure of uncertainty for both linear and
curvilinear equations has specific merit as it can be readily used in the value chain
models.

There are limits to this approach of incorporating uncertainty in process modelling,

especially where the response is a function of more than one input variable.
Reducing the dimensionality of relationships between primary and response
variables using multiple or partial regression models (e.g., Nipals algorithm
(Esbensen, 2002)) is one alternative solution. Another approach does not reduce
the input data dimensionality but aims to model the response variable using
multiple multidimensional response surfaces. This requires that sampling
experiments be designed to collect multivariate test data that are isotopic (i.e. the
same data types collected for all samples from all locations) is collected at the right
scale. This data makes it possible to generate response surfaces using some
interpolation method - such as kriging or radial basis functions. These modelled
surfaces can then be perturbed using the error of the fit to represent uncertainty
and then be applied to the value chain model. This development is currently being
embedded in the value chain system that has resulted from this research and is a far
more rapid method of deriving multiple responses from estimated block
characteristics.

232 Stephen Coward - April 2020

 Chapter 10: Summary and Conclusions

The data acquisition experiment required several different types of data to be

acquired from several different laboratories. The co-ordination of the sample
movements was not trivial and is perhaps one of the major hurdles that can be
overcome in acquiring valid and reliable rock characteristic data. Where practical it
is recommended that part of the trade-off studies that should be carried out when
planning sampling designed to acquire data for recovery factor estimation includes
the accessibility of more than one type of measurement at each laboratory. This aids
substantially in tracking the sample and data. The information acquired from
downhole geophysics has shown itself to be of substantial use not only for
estimating physical rock characteristics, but also for providing a basis for improving
the location of the acquired core.

10.5 Recommended areas for additional research

Several of the findings of this research have resulting in suggestions for either a
deeper investigation or additional experimentation to validate and expand specific
findings.

In tandem with this research, the author has been responsible for adapting the
methodology developed here to quantify recovery in various operations and some
of these models have been reported in the literature. This has resulted in the
identification of areas which required additional work to enable value chain models
to be applied to projects for other mined commodities including iron ore, uranium
and gold projects.

Evolution of access to less costly computing power continues to enable a far greater
range of more complex unit process models. This includes access to online cloud
infrastructure that allows access to multiple configurable servers for limited time at
almost no cost. The value chain approach lends itself to parallel computing. A single
orebody, single mine plan, single process configuration and associated financial
model constitutes a single yet complete value chain model, and so it is possible to
distribute each value chain model to disparate servers for processing. Work is
continuing in this area and is likely to lead to a reduction of several orders of
magnitude in processing speed.

Several of the non- destructive geophysical tools were identified as having some
weak yet discernible correlation with the destructive measures. As these tools
improve and the ratios of the signal strength to noise increases the use for this
additional, relatively cheap, data will increase. This will require specific calibration
of the tools to the environments and mineralogy of the rocks in which they are to be
used.

Stephen Coward - April 2020 233

 Evaluation of metallurgical recovery factors for diamonds recovered from kimberlites

Value chain model validation can be improved by using this approach on existing
operations in a predictive way. Ensuring that the value chain model that includes
orebody characteristics and a suitable process model is up to date can provide
several benefits to these operations. This includes being able to create a production
performance forecast, and to identifying periods in which the metallurgical recovery
factor will be compromised because of variable rock characteristics. This approach
will allow mine decision-makers to take pre-emptive action to mitigate recovery
risk and harness opportunities for increasing recovery and process performance.

234 Stephen Coward - April 2020

 Chapter 11: References

11 REFERENCES

Agricola, G., 1556. De Re Metallica. 1950 ed., New York: Dover Publications.

Aitchison, J. & Brown, J.A.C., 1957. The Lognormal Distribution, Cambridge,

England: Cambridge University Press.

Anglo American, 2018. Anglo American Ore Reserves and Mineral Resources Report
2018., p.53.

Ameluxen, P. 2003. The application of the SAG Power Index to orebody hardness
characterization for the design and optimization of comminution circuits. McGill
University, Montreal, Canada.

American Society for Testing and Materials,1988: Standard Method for Splitting
Tensile Strength of Intact Rock Core Specimens D 3967-86. 1988 Annual Book of
ASTM Standards, Vol. 04.08, Soil and Rock; Building Stones; Geotextiles, 471–475.

American Society for Testing and Materials 1998: D7012 - 14e1 Standard Test
Methods for Compressive Strength and Elastic Moduli of Intact Rock Core
Specimens under Varying States of Stress and Temperatures.

Andersen, J.S. 1988. Development of a cone crusher model. M.Eng.Sc Thesis,

University of Queensland, Brisbane.

Appleyard, G.R. ,2001. An Overview and Outline, in Mineral resource and Ore
Reserve estimation- The AusIMM guide to good practice’, in Edwards, A. . (ed.).
Melbourne: The Australasian Institute of Mining and Metallurgy, pp. 3–12.

Armstrong, M. 1988. Geostatistics volume 1 and 2, 268 p (Kluwer Academic

publishers' group: Amsterdam, The Netherlands).

Ashley, K.J. and Callow, M.I. 2000. Variability: Exercises in Geometallurgy [online],
Available from: <http://e-j.com/ar/mining_ore_variability_exercises/; > [Accessed:
10 March 2004].

Aspinall, W.P., 2018. Structured Elicitation of Expert Judgment for Probabilistic

Hazard and Risk Assessment in Volcanic Eruptions. In Statistics in Volcanology. The

Stephen Coward - April 2020 235

 Evaluation of metallurgical recovery factors for diamonds recovered from kimberlites

Geological Society of London on behalf of The International Association of

Volcanology and Chemistry of the Earth’s Interior, pp. 15–30.

American Society for Testing and Materials, 1998. ASTM Standards on

Environmental Sampling.

Bagnell, W., Bedell, P., Bertrand, V., Brummer, R., Farrow, D., Gagnon, C.L.G., Gormely,
L., Magnan, M. and St-Onge, J. 2013. NI 43-101 Technical Report for The Renard
Diamond Project, Québec, Canada, Stornoway Diamond Corporation, Québec,
Canada.

Bearman, R.A., Pine, R.J., and Wills, B.A. 1989. Use of Fracture Toughness Testing in
Characterizing the Comminution Potential of Rock, Proceedings of MMIJ/IMM Joint
Symposium, Kyoto pp161-170

Beniscelli, J., Carrasco, P., Dowd, P.A. Ferguson, G. and Tulcanaza, E. 2000. Estimation
of Resources and Conversion to Reserves - Protocols for the Assessment, Reduction
and Management of Risk, paper presented to Mass Min 2000, Brisbane, Australia.

Bojcevski, D., Vink, L., Johnson, N.W., Landmark, V., Johnston, M., Mackenzie, J. and
Young, M. F. 1998. Metallurgical characterisation of George Fisher Ore Textures and
Implications for Ore Processing, paper presented to Mine to Mill conference,
Brisbane Australia.

Boulding, K.E., 1956. General Systems Theory - The Skeleton of Science. Management
Science, 2(3), pp.197-208

Box, G. E. P. and Draper, N. R. 1987. Empirical Model-Building and Response Surfaces

John Wiley & Sons, Inc.

Boychuk, K.G., Garcia, D. H., Sharp, A.W., Vincent, J.D. and Yeomans, T. J., 2012. NI 43-
101 Technical Report on the Pitarrilla Project, Durango State, Mexico, Silver
Standard Resources.

Bradfield, R., Wright, G., Burt, G., Cairns, G., Van Der Heijden, K., 2005. The origins
and evolution of scenario techniques in long range business planning. Futures, 37(8),
pp.795–812.

Bratvold, B. and Begg, S. 2002. Would You Know a Good Decision if You Saw One,
paper presented to SPE Annual Technical Conference, San Antonio, Texas.

Brennan, M.J. and Schwartz, E.S. 1985. Evaluating Natural Resource Investments,
Journal of Business, 58(2).

236 Stephen Coward - April 2020

 Chapter 11: References

Brown, R., Tait, M., Field, M., Sparks, R.S.J. 2008. Geology of a complex kimberlite
pipe (K2 pipe, Venetia Mine, South Africa): Insights into conduit processes during
explosive ultrabasic eruptions. Bulletin of Volcanology, 71, pp.95–112.

Burbeck, S. 1992. Applications Programming, how to use Model-View-Controller

(MVC). Applications Programming in Smalltalk-80, p.3. Available at:
https://web.archive.org/web/20120729161926/http://st-
www.cs.illinois.edu/users/smarch/st-docs/mvc.html [Accessed September 14, 2011].

Bush, E.D. 2010. An overview of the estimation of kimberlite diamond deposits.

Southern African Institute of Mining and Metallurgy: Diamonds - Source to Use 2010,
pp.73–84.

Bye, A.R. 2011. Case Studies Demonstrating Value from Geometallurgy Initiatives. In
Geomet 2011. pp. 5–7.

Caers, J. and Rombouts, L. 1996. Valuation of Primary Diamond Deposits by Extreme

Vale Statistics, Economic Geology 91( ):841-854.

Campbell, J.A.H. 2009. THE DEVELOPMENT OF AK6. In Diamonds Source to Use

2009. Gaberone: The Southern African Institute of Mining and Metallurgy, pp. 1–20.

Carrasco, P., Chiles, J. P., Séguret, S.A. 2008. Additivity, Metallurgical Recovery, and
Grade, in Geostats 2008: VIII International Geostatistics Congress (eds: J Ortiz and X
Emery), Santiago, December, pp 465-476.

Chapman, J. G. and Boxer, G. L. 2003. Size distribution analyses for estimating

diamond grade and value, paper presented to 8th International Kimberlite
Conference, Victoria.

Cookenboo, H.O. & Grutter, H.S. 2010. Mantle-derived indicator mineral

compositions as applied to diamond exploration. Geochemistry: Exploration,
Environment, Analysis, 10, pp.81–95.

Copur, H., Billgin, N., Tuncdemir, H. and Balci, C. 2003. A set of indices based on
indentation tests for assessment of rock cutting performance and rock properties,
South African Institute of Mining and Metallurgy.

Cornah, A., Vann, J., & Driver, I. 2013. Comparison of three geostatistical approaches
to quantify the impact of drill spacing on resource confidence for a coal seam (with
a case example from Moranbah North, Queensland, Australia). International Journal
of Coal Geology, 112, 114-124. DOI: 10.1016/j.coal.2012.11.00

Coward, S.J. and Ferreira, J. 2005. Development and conceptual evaluation of a

methodology for sampling for diamonds in broken ore. In EMMES Conference. p. 149.

Stephen Coward - April 2020 237

 Evaluation of metallurgical recovery factors for diamonds recovered from kimberlites

Coward, S.J., Vann, J., Dunham. S., Stewart, M. 2009. The Primary-Response
Framework for Geometallurgical Variables. In Seventh International Mining Geology
Conference. pp. 109–113.

Daniel, M., Lane, G. and McLean, E. 2010. Efficiency, economics, energy and
emissions – Emerging criteria for comminution circuit decision making., paper
presented to XXV International Mineral Processing Congress (IMPC) Brisbane, Qld,
Australia, 6 - 10 September 2010.

Davis, G.A. 1995. An Investigation of the Under Pricing Inherent in DCF Valuation
Techniques, paper presented to SME Annual Meeting, Denver, Colorado, 6-9th
March 1995.

De Beers Group Services, 2017. De Beers Annual Finance Seminar, London UK.

De Beers Mineral Resource Intelligence Unit, 2005. Estimates of Global Diamond

Production.

Demming, W.E. 1986. Out of the Crisis, Massachusetts Institute of Technology,

Centre for Advanced Engineering Study.

Dimitrakopoulos, R. and Ramazan, S. 2004. Uncertainty Based Production

Scheduling in Open-Pit Mining, SME Transaction, 316(03-151): pp 1-9.

Lamghari, A. & Dimitrakopoulos, R., 2016. Network-flow based algorithms for

scheduling production in multi-processor open-pit mines accounting for metal
uncertainty. European Journal of Operational Research, 250(1).

Dowd, P.A. 1978. Advances in Geostatistics: Numerical Methods and their

Application. Ph.D. Thesis, University of Leeds, UK

Dowd, P. A. 1992. A review of recent developments in Geostatistics. Computers and

Geosciences, 17(10): pp. 1481-1500.

Dowd, P. A. 1994, 2004. Geostatistical Simulation Course Notes.

Dowd, P.A. 1997. Geostatistical Characterisation of Three-Dimensional Spatial

Heterogeneity of Rock Properties at Sellafield. Transactions of the Institution of
Mining and Metallurgy, 106, A133-197. London, United Kingdom.

Dowd, P.A. and Dare-Bryan, P.C, 2004. Planning Designing and Optimising
Production Using Geostatistical Simulation. Proceedings of the International

238 Stephen Coward - April 2020

 Chapter 11: References

Symposium on Orebody Modelling and Strategic Mine Planning. Pub AusIMM

(Melbourne). ISBN 1 920806 22 9; 321-338.

Dowd, P.A. and Pardo-Igúzquiza, E., 2006. Core-log integration: optimal

geostatistical signal reconstruction from secondary information, Transactions of the
Institution of Mining and Metallurgy, Section B, 115(2):12.

Esbensen, K. 2002. Multivariate Data Analysis - In Practice, 5th ed, 598 p (Camo
Process AS: Oslo).

Esbensen, K. 2004. 50 Years of Pierre Gy's "Theory of Sampling", Chemometrics and

Intelligent Laboratory systems, 74(1): pp3-6.

Farrow, D.J. 2015. 2015 Mineral Resource Update for the Renard Diamond Project -
Ni 43-101 Technical Report, Vancouver.

Ferreira, J.J., 2013. Sampling and estimation of diamond content in kimberlite

based on microdiamonds. PhD thesis, l’École nationale supérieure des mines de
Paris, Paris.

Field, M., Stiefenhofer, J., Robey, J. and Kurszlaukis, S. 2008. Kimberlite-hosted

diamond deposits of southern Africa: A review., Ore Geology Reviews, 34:33-75.

Field, M. and Scott Smith, B.H. 1999, ‘Contrasting geology and near-surface
emplacement of kimberlite pipes in Southern Africa and Canada’, The J. B. Dawson
Volume, pp. 214–237.

Fullagar, P.K. and Fallon, G.N. 1997. Orebody Delineation and Rock Mass
Characterisation, paper presented to Explor97, Toronto, Canada, 14-18 September
1997.

Garman, M.B. and Kohlhagen, S.W. 1983. Foreign currency option values,
International Money Finance, Vol 2: p231-237.

Goldratt, E.M. 1990. Theory of Constraints. Great Barrington: North River Press, pp.1–
159.

Griffith, A.A. 1921. The Phenomenon of Rupture and Flow in Solids, Philosophical
Transactions of the Royal Physical Society, Vol. 221, pp. 163-198.

Gy, P. 2004. Sampling of Discrete Materials - A New Introduction to the Theory of

Sampling, Chemometrics and Intelligent Laboratory Systems, 74(1): pp 7-24.

Stephen Coward - April 2020 239

 Evaluation of metallurgical recovery factors for diamonds recovered from kimberlites

Hawthorne, J.B. 1975. Model of a kimberlite pipe. In First International Kimberlite

Conference. Physics and Chemistry of the Earth, v. 9, pp. 1–16.

Holmes, J.A. 1957. Contribution to study of comminution -- Modified form of Kick's

law. Institution of Chemical Engineers -- Transactions, 35(2): 125-141.

ISRM, 1972. Suggested Methods for Determining the Uniaxial Compressive Strength
and Deformability of Rock Materials.

ISRM, 1994. Suggested Methods for Determining Mode 1 Fracture Toughness Using
Cracked Chevron Notched Brazilian Disc.

Johnson, D., Meikle, K., Pilotto, D. and Lone, K. 2014. NI 43-101 Technical Report for
the Gahcho Kué Project 2014 Feasibility Study Mountain Province Diamonds Inc.,
Montréal.

JORC, 2012. Australasian Code for Reporting of Exploration Results, Mineral

Resources and Ore Reserves (The JORC Code) ) [online]. Available from:
<http://www.jorc.org> (The Joint Ore Reserves Committee of The Australasian
Institute of Mining and Metallurgy, Australian Institute of Geoscientists and
Minerals Council of Australia).

Journel, A.G. 1974. Simulations Conditionnelles - Theorie et Practique, Centre de

Geostatistique, Fontainebleau, France.

Journel, A.G. and Huijbrechts, C. 1978. Mining Geostatistics, 600 p (Academic Press:
London).

Kahneman, D., 2011. Thinking fast, thinking slow 1st ed., New York: Farrar, Straus and
Giroux.

Keeney, L. and Walters, S. 2008. Geometallurgical modelling. GeMIII (Amira P843)

Technical Report I, p 16.1 – 16.24.

King, R.P. 2001. Modelling and Simulation of Mineral Processing Systems,

(Butterworth-Heinemann.: Oxford).

Kjaarsgaard, B.A. 1995. Geological of Canadian Mineral Deposit types, p560 p

(Geological Survey of Canada: Ottawa).

Kleingeld, W.J. 1976. Reconciliation of Ore Reserves in the Number One Diamond
Mining Area.

240 Stephen Coward - April 2020

 Chapter 11: References

Kleingeld, W. 1982. A reconciliation of diamond losses at CDM, Internal Report, De

Beers Consolidated Mines, South Africa.

Kleingeld, W. J., Lantuéjoul, C., Prins, C. F. and Thurston, M.L. 1996. The Conditional
Simulation of a Cox Process with Application to Diamond Deposits and Discrete
Particles, paper presented to Geostatistics Congress, Wollongong, Australia.

Kleingeld, W.J. and Nicholas, G.D. 2004. Diamond Resources and Reserves -
Technical Uncertainties affecting their Estimation, Classification and Evaluation.
Proceedings of the International Symposium on Orebody Modelling and Strategic
Mine Planning. Pub AusIMM (Melbourne). ISBN 1 920806 22 9;

Kolmogorov, A.N. 1941. Uber das logarithmisch normale Verteilungsgesetz der

Dimensionen der Teilchen bei Zerst¨uckelung. Dokl. Akad. Nauk. SSSR 31, 99–
101.Translated as “The logarithmically normal law of distribution of dimensions of
particles when broken into small parts”, NASA technical translation NASA TTF–
12,287, NASA, Washington, D. C., June 1969

Krige, D.G., Guarascio, M. & Camisani-Calzolari, F.A. 1988. Early South African
Geostatistical Techniques in Todays’s Perspective. In M. Armstrong, ED.
Geostatistics: Proceedings of the Third International Geostatistics Congress.
Avignon, p. 1027.

Lane, K.F. 1988. The economic definition of ore: cut-off grades in theory and practice,
London: Mining Journal Books.

Lantuéjoul, C. 1990. The Benefit Function, Internal Report, De Beers Consolidated

Mines Pty Ltd., Johannesburg, South Africa.

Lantuéjoul, C. 1998. On liberation and breakage, Internal Report, De Beers

Consolidated Mines Pty Ltd., Johannesburg, South Africa.

Lerchs, H. and Grossman, I.F., 1965. Optimum Design of Open Pit Mines,
Transactions CIM, 68(633): pp17-24.

Lewis, P. J., 2001. Metallurgical Evaluation Leading to the Determination of Ore

Reserves, in Mineral Resource and Ore Reserve Estimation, pp359–368 p (The
Australasian Institute of Mining and Metallurgy: Melbourne).

Lilford, B. 2003. Design Criteria for Voorspoed, De Beers, Johannesburg, South

Africa.1

Lilford, E.V. & Minnitt, R.C.A. 2005. A comparative study of valuation methodologies
for mineral developments. Southern African Institute of Mining and Metallurgy,
pp.p29-41.

Stephen Coward - April 2020 241

 Evaluation of metallurgical recovery factors for diamonds recovered from kimberlites

Lynn, M., Nowicki, T., Valenta, M., Robinson, B., Gallagher, M., Bolton, R. and Sexton,
J. 2014. Karowe Diamond Mine, Botswana, NI 43-101 Independent Technical Report
Lucara Diamond Corp., Vancouver, BC, Canada.

Mackey, P.J. and Nesset, J.E. 2003. The impact of commissioning and start-up
performance on a mining/metallurgical project, paper presented to Proceedings of
the 35th Annual Meeting of the Canadian Mineral Processors.

Macnae, J. 1995. Applications of geophysics for the detection and exploration of

kimberlites and lamproites. Journal of Geochemical Exploration, 53(1-3), pp.213-
243.

Markowitz, H. M. 1952. Portfolio Selection, Journal of Finance, Vol. 7: p77-91.

Matheron, G. 1963. Principles of Geostatistics. Economic Geology, 58(8), pp.1246–

1266.

Matheron, G. 1973. The Intrinsic Random Functions and their Application, Advances
in Applied Probability, 5: pp 439-468.

McKee, D.J., Chitombo, G.P., & Morrell, S. 1995. The relationship between
fragmentation in mining and comminution circuit throughput. Minerals
Engineering, 8(11), 1265–1274. https://doi.org/10.1016/0892-6875(95)00094-7

McBean, D., Kirkley, M. and Revering, C. 2001. Structural controls on the morphology
of the Snap Lake Dyke. 8th International Kimberlite Conference.

Moore, G.E. 1965. Cramming More Components onto Integrated Circuits,

Electronics, April 19, 1965. Electronics, 38(8), pp.82–85.

Momayez, M., Sadri, A. & Hassani, F.P. 1995. MSR: A technique for determining the
mechanical properties of rocks. In 35th US Rock Mechanics Symposium. Lake Tahoe,
Nevada, pp. 843–848.

Morrell, S., Dunne, R. C. and Finch, W. 1993. The liberation of a grinding circuit
treating gold bearing ore, paper presented to XVIII Int. Min Proc Cong, Sydney.

Morell, S. 2003. Predicting the Specific Energy of Autogenous and Semi Autogenous
Mills from Small Diameter Drill Core Samples Minerals Engineering, 17: pp 447-451.

Morkel, J. 2006. Kimberlite weathering: mineralogy and mechanism, University of

Pretoria, Pretoria.

242 Stephen Coward - April 2020

 Chapter 11: References

Napier-Munn, T., Morell, S., Morison, R. and Kojovic, T. 1999. Mineral Comminution
Circuits, Their Operation and Optimisation, 2nd ed, 413 p (JKMRC, University of
Queensland: Indooroopilly).

NI43-101, 2001. National Instrument 43-101, Standards of Disclosure for Mineral

Projects, Canadian Institute of Mining, 21.

Nicholas, G., Coward, S., Armstrong, M. and Galli, A. 2006. Integrated Mine Evaluation
- Implications for Mine Management. In: AUSIMM. ed. International Mine
Management Conference. Melbourne. Australia.

Nicholas, G.D., Coward, S. J., Rendall, M., and Thurston, M.L. 2007. Decision-Making
Using an Integrated Evaluation Model Versus Sensitivity Analysis and Monte Carlo
Simulation. In Canadian Institute of Mining and Metallurgy International conference.
Montreal, Canada.

Nicholas, G., Coward, S. & Ferreira, J. 2008. Financial Risk Assessment Using
Conditional Simulations in an Integrated Evaluation Model. In the Eighth
International Geostatistics Congress. Santiago, Chile, p. 10.

Olea, R.A. 1991. Geostatistical Glossary and Multilingual Dictionary, Oxford

University Press, New York.

Olsson, W. 1974. Grain Size Dependence of Yield Stress in Marble, Journal of

Geophysical Research, (November 1974): pp 4859-4862.

Ozturk, C. N. E., Bilgin, N. 2004. The assessment of Rock cut ability, and physical and
mechanical rock properties from a texture coefficient, Journal of the South African
Institute of Mining and Metallurgy.

Parker, H. 1997. Applications of Geostatistical Methods in Exploration Programme

Design., National Council for United States - China Trade Technical Exchange,
Peking.

Petersen, K., 2005. Development of a Flowsheet Model for the Snap Lake Mine.
Unpublished Report.

Petra Diamonds, 2019. Petra Diamonds Limited Annual Report 2019, London UK.

Stephen Coward - April 2020 243

 Evaluation of metallurgical recovery factors for diamonds recovered from kimberlites

Plitt, L. R. 1976. A Mathematical Modelling of the Hydro cyclone Classifier, CIM

Bulletin 69(776):114-123.

Rappaport, 2009. Rapaport price list - July-03., available at

http://store.rapaport.com, accessed 2009/03/07.

Richmond, A.J. 2003. Multi-Scale Ore Texture Modelling for Mining Applications,
University of Brisbane, Brisbane.

Rittinger, Peter Ritter von. 1867. Lehrbuch der Aufbereitungskunde, Berlin: Verlag
von Ernst & Korn.

Robinson, G.K. 2004. Managing variation of bulk materials in stockpiles and

mineral processing plants. Chemometrics and Intelligent Laboratory Systems, 74,
pp.121–133.

Rombouts, L. 1995. Sampling and statistical evaluation of diamond deposits, Journal

of Geochemical Exploration, 53:351-367.

Royle, A.G. 1986. Alluvial Sampling Formula and Recent Advances in Alluvial Deposit
Evaluation, Transactions of The Institution of Mining and Metallurgy: pp B179-
B182.

Samis, M., Laughton, D., and Davis, G.A. 2006. Valuing uncertain asset cash flows
when there are no options: A real options approach, Resources Policy, 30: p285-298.

SAMREC, 2016. The South African Code for the Reporting of Exploration Results,
Mineral Resources, and Mineral Reserves (the SAMREC code).

SAMREC, 2017. SAMREC Guideline Document for the Reporting of Diamond

Exploration Results, Diamond Resources and Diamond Reserves.

Shipiki-Kali, M. 2000. A Comparative analysis of Operational Costs - 2000. Internal

Report, De Beers, Johannesburg, South Africa.

Silva, D.S.F. & Boisvert, J.B. 2014. Mineral resource classification: a comparison of
new and existing techniques. Journal of the Southern African Institute of Mining and
Metallurgy, 114, pp.265–273. Available at:
http://www.scielo.org.za/scielo.php?script=sci_arttext&pid=S2225-
62532014000300017&nrm=iso.

244 Stephen Coward - April 2020

 Chapter 11: References

Skinner, E.M.W. and Clement, C.R. 1979. Mineralogical classification of southern

African kimberlites. In F. R. Meyer & H. O. A. Boyd, EDS. Proceedings of 2nd
International Kimberlite Conference. Washington, pp. 129–139.

Sothcott, J., Hennah, S.J., Mc Cann, C., Black, S and Stevenson, I. 2005. Measurement
of the Broadband Acoustic Properties of Kimberlite: A Feasibility Study, De Beers
Consolidated Mines South Africa, Johannesburg.

South African Institute for Mining and Metallurgy, 2002. The SAMVAL code, Draft
Standards and guidelines for valuation of mineral projects, properties and assets in
the mining industry of South Africa.

Sparks, R.S.J. 2006. Dynamical constraints on kimberlite volcanism. Journal of

Volcanology and Geothermal Research, 155(1–2), pp.18–48.

Sparks, R.S.J., Brooker, R., Field, M. 2009. The nature of erupting kimberlite melts.
Lithos, 112, pp.429–438.

Sperra, F.J. 2000. Encyclopaedia of Volcanoes, London: Academic Press.

Taggart, A.F. 1964. Handbook of Mineral Dressing, 8th ed (John Wiley and Sons:
London).

Tavares, L. & King, R.P. 1998. Single-particle fracture under impact loading.
International Journal of Mineral Processing, 54, pp.1–28.

De Beers Consolidated Mines Limited , 2001., Technical and Financial Report, 149
p.

Thurston, M.L. 1998. The quantification of uncertainty attached to selected sampling

protocols in a kimberlite using a discrete simulation method, University of the
Witwatersrand, Johannesburg.

Vose, D. 2002. Risk Analysis, a Quantitative Guide (John Wiley and Sons Ltd:
London).

Wackernagel, H. 1995. Multivariate Geostatistics, 256 p (Springer-Verlag: Berlin).

Whiten, W.J. 1972. Simulation and Model Building for Mineral Processing, PhD
dissertation - University of Queensland, Australia.

Stephen Coward - April 2020 245

 Evaluation of metallurgical recovery factors for diamonds recovered from kimberlites

12 APPENDICES

The appendices provided here are to give the interested reader more technical
substantiation of aspects of metallurgical recovery estimation that are addressed by
this research.

246 Stephen Coward - April 2020

 Chapter 12: Appendices

APPENDIX 1 -EXPERIMENTAL DESIGN AND DATA FROM OREBODY SAMPLING

Brief depiction of the layout, the types of samples collected and the raw data that
were generated from this experiment. In Figure 71 a surface schematic is shown
depicting the area in which the sample holes were located.

Figure 71: Schematic showing the location of samples from K2 pipe.

Two categories of drill holes were defined. The first comprises holes where all the
material was sampled, and the second where only portions were sampled. In Figure
72 the layout of the holes is shown with the different coloured banding representing
the type and location of samples. The white areas of cores were not sampled. The
full details of the sampling are contained in Figure 73 and in Figure 74

Stephen Coward - April 2020 247

 Evaluation of metallurgical recovery factors for diamonds recovered from kimberlites

Figure 72: A view of the layout of the core holes depicting the location of the subsamples

The sampling layout of the cores extracted is depicted in Figure 72.White

intersections were not sampled but retained; all other coloured intersections were
sampled. Details are of the subsamples are shown in Figure 73 and Figure 74.

248 Stephen Coward - April 2020

 Chapter 12: Appendices

Log document name Geomet Analysis and Scans

Drill hole name GM-1E
Scale 1 / 150
Depth portion 0.00 M - 50.00 M
Printing date 23/09/04

TO (M) A_CODE TO (M) ANALYSIS SAMPNO DMT SPECTRAL GEOPHYSI

0.0 Pet GM-1E/0m/Pet
0.5 Mida GM-1E/0.5m/Mida
1.8 Pet GM-1E/1.75m/Pet
2.0 Geomet GM-1E/2m/Geomet
3.0 Geotech GM-1E/3m/Geotech
3.5 Mida GM-1E/3.5m/Mida

5.0 4.8 GM-1E/4.75m/Pet

Pet
5.0 GM-1E/5m/Pet
5.5 Mida GM-1E/5.5m/Mida
6.8 Pet GM-1E/6.75m/Pet
7.0 Geomet GM-1E/7m/Geomet
8.0 Geotech GM-1E/8m/Geotech
8.5 Mida GM-1E/8.5m/Mida

10.0 9.8 GM-1E/9.75m/Pet

Pet
10.0 GM-1E/10m/Pet
10.5 Mida GM-1E/10.5m/Mida
11.8 Pet GM-1E/11.75m/Pet
12.0 Geomet GM-1E/12m/Geomet
13.0 Geotech GM-1E/13m/Geotech
13.5 Mida GM-1E/13.5m/Mida

15.0 14.8 GM-1E/14.75m/Pet

Pet
15.0 GM-1E/15m/Pet
15.5 Mida GM-1E/15.5m/Mida
16.8 Pet GM-1E/16.75m/Pet
17.0 Geomet GM-1E/17m/Geomet
18.0 Geotech GM-1E/18m/Geotech
18.5 Mida GM-1E/18.5m/Mida

20.0 19.8 GM-1E/19.75m/Pet

Pet
20.0 GM-1E/20m/Pet
20.5 Mida GM-1E/20.5m/Mida
21.8 Pet GM-1E/21.75m/Pet
22.0 Geomet GM-1E/22m/Geomet
23.0 Geotech GM-1E/23m/Geotech
23.5 Mida GM-1E/23.5m/Mida

25.0 24.8 GM-1E/24.75m/Pet

Pet
25.0 GM-1E/25m/Pet
25.5 Mida GM-1E/25.5m/Mida
26.8 Pet GM-1E/26.75m/Pet
27.0 Geomet GM-1E/27m/Geomet
28.0 Geotech GM-1E/28m/Geotech
28.5 Mida GM-1E/28.5m/Mida

30.0 29.8 GM-1E/29.75m/Pet

Pet
30.0 GM-1E/30m/Pet
30.5 Mida GM-1E/30.5m/Mida
31.8 Pet GM-1E/31.75m/Pet
32.0 Geomet GM-1E/32m/Geomet
33.0 Geotech GM-1E/33m/Geotech
33.5 Mida GM-1E/33.5m/Mida

35.0 34.8 GM-1E/34.75m/Pet

Pet
35.0 GM-1E/35m/Pet
35.5 Mida GM-1E/35.5m/Mida
36.8 Pet GM-1E/36.75m/Pet
37.0 Geomet GM-1E/37m/Geomet
38.0 Geotech GM-1E/38m/Geotech
38.5 Mida GM-1E/38.5m/Mida

40.0 39.8 GM-1E/39.75m/Pet

Pet
40.0 GM-1E/40m/Pet
40.5 Mida GM-1E/40.5m/Mida
41.8 Pet GM-1E/41.75m/Pet
42.0 Geomet GM-1E/42m/Geomet
43.0 Geotech GM-1E/43m/Geotech
43.5 Mida GM-1E/43.5m/Mida

45.0 44.8 GM-1E/44.75m/Pet

Pet
45.0 GM-1E/45m/Pet
45.5 Mida GM-1E/45.5m/Mida
46.8 Pet GM-1E/46.75m/Pet
47.0 Geomet GM-1E/47m/Geomet
48.0 Geotech GM-1E/48m/Geotech
48.5 Mida GM-1E/48.5m/Mida
49.8 Pet GM-1E/49.75m/Pet
50.0

Figure 73:Listing of subsamples taken from cores that were fully sampled

Stephen Coward - April 2020 249

 Evaluation of metallurgical recovery factors for diamonds recovered from kimberlites

Log document name Geomet Analysis and Scans

Drill hole name GM-2E
Scale 1 / 150
Depth portion 0.00 M - 50.00 M
Printing date 23/09/04

TO (M) A_CODE TO (M) ANALYSIS SAMPNO DMT SPECTRAL GEOPHYSI

0.0 Retain GM-2E/0m/Retain

2.5 Pet GM-2E/2.5m/Pet

3.0 Mida GM-2E/3m/Mida
4.3 Pet GM-2E/4.25m/Pet
5.0 4.5 Geomet GM-2E/4.5m/Geomet
5.5 Geotech GM-2E/5.5m/Geotech
6.0 Mida GM-2E/6m/Mida
7.3 Pet GM-2E/7.25m/Pet

7.5 GM-2E/7.5m/Retain

10.0 Retain

10.0 GM-2E/10m/Retain

12.5 Pet GM-2E/12.5m/Pet

13.0 Mida GM-2E/13m/Mida
14.3 Pet GM-2E/14.25m/Pet
15.0 14.5 Geomet GM-2E/14.5m/Geomet
15.5 Geotech GM-2E/15.5m/Geotech
16.0 Mida GM-2E/16m/Mida
17.3 Pet GM-2E/17.25m/Pet

17.5 GM-2E/17.5m/Retain

20.0 Retain

20.0 GM-2E/20m/Retain

22.5 Pet GM-2E/22.5m/Pet

23.0 Mida GM-2E/23m/Mida
24.3 Pet GM-2E/24.25m/Pet
25.0 24.5 Geomet GM-2E/24.5m/Geomet
25.5 Geotech GM-2E/25.5m/Geotech
26.0 Mida GM-2E/26m/Mida
27.3 Pet GM-2E/27.25m/Pet

27.5 GM-2E/27.5m/Retain

30.0 Retain

30.0 GM-2E/30m/Retain

32.5 Pet GM-2E/32.5m/Pet

33.0 Mida GM-2E/33m/Mida
34.3 Pet GM-2E/34.25m/Pet
35.0 34.5 Geomet GM-2E/34.5m/Geomet
35.5 Geotech GM-2E/35.5m/Geotech
36.0 Mida GM-2E/36m/Mida
37.3 Pet GM-2E/37.25m/Pet

37.5 GM-2E/37.5m/Retain

40.0 Retain

40.0 GM-2E/40m/Retain

42.5 Pet GM-2E/42.5m/Pet

43.0 Mida GM-2E/43m/Mida
44.3 Pet GM-2E/44.25m/Pet
45.0 44.5 Geomet GM-2E/44.5m/Geomet
45.5 Geotech GM-2E/45.5m/Geotech
46.0 Mida GM-2E/46m/Mida
47.3 Pet GM-2E/47.25m/Pet

47.5 Retain GM-2E/47.5m/Retain

50.0

Figure 74: Depiction of the subsamples taken from cores that were partially sampled.

250 Stephen Coward - April 2020

 Chapter 12: Appendices

APPENDIX 2 -OREBODY CHARACTERISTIC ESTIMATES

The data gathered during the orebody sampling experiment were used in several
ways to generate spatial estimates of the primary and response characteristics of
the orebody. This Appendix presents images of the data obtained from the Venetia
K2 orebody.

Figure 75:A view from the south-west of Venetia K2 sampled area - showing downhole
density.

Stephen Coward - April 2020 251

 Evaluation of metallurgical recovery factors for diamonds recovered from kimberlites

Figure 76:A view from the south-west of Venetia K2 sampled area - showing downhole
density and P wave velocity.

Figure 77:A view from the south-west of Venetia K2 sampled area - UCS sample values as
scaled spheres and estimated block density in transparent blocks.

252 Stephen Coward - April 2020

 Chapter 12: Appendices

Figure 78:A view from the south-west of Venetia K2 sampled area - t10 sample values as
spheres and estimated t10 in transparent blocks.

Stephen Coward - April 2020 253