Advances in Orebody Modelling and Strategic Mine Planning I PDF

Advances in
Orebody Modelling and

Strategic Mine Planning I
Old and New Dimensions in a Changing World
Editor: R Dimitrakopoulos
Number 17
Advances in
Orebody Modelling and
Strategic Mine Planning I
Old and New Dimensions in a Changing World
Edited by
Roussos Dimitrakopoulos
The Australasian Institute of Mining and Metallurgy

Spectrum Series No 17
COSMO
Stochastic Mine Planning Laboratory
Published by
THE AUSTRALASIAN INSTITUTE OF MINING AND METALLURGY
Level 3, 15 - 31 Pelham Street, Carlton Victoria 3053 Australia
© The Australasian Institute of Mining and Metallurgy 2010
All papers published in this volume were refereed prior to publication.
The Institute is not responsible as a body for the facts and opinions advanced in any of its publications.
ISBN 978 1 921522 21 5
Desktop published by:

Kylie McShane and Olivia Tet Fong for
The Australasian Institute of Mining and Metallurgy
Compiled on CD ROM by:

Visual Image Processing Pty Ltd
PO Box 3180
Doncaster East VIC 3109
Reviewers
We would like to thank the following people for their contribution
towards enhancing the quality of papers included in this volume.
Enrique Acuna Ute Mueller
Jorg Benndorf Hussein Mustapha
Alexandre Boucher Mike O‘Brien
Marcus Brazil Harry Parker
Peter Carter Richard Peattie
Snehamoy Chatterjee Salih Ramazan
Michel Dagbert Jean-Michel Rendu
Roussos Dimitrakopoulos Andrew Richmond
Tamer Dincer Sabry Abdel Sabour
Peter Dowd Michael Samis
Marcelo Godoy Peter Stone
Nicole Grieco Olivier Tavchandjian
John Hadjigeorgiou Erkan Topal
Geoff Hall Marcus Volz
Xu Huan Chris Wharton
Arja Jewbali David Whittle
Brett King Jeff Whittle
Amina Lamghari Gavin Yeates
Conor Meagher
Foreword
In an ever-changing world of uncertain markets, mismatches between demand and reserve base growth
(supply), the increasingly critical issue of sustainable development, emerging technologies and new
technical problems and their solutions, this book represents a contribution to both old and new
dimensions of a most intricate, demanding and fundamentally important technical aspect of mining
ventures and the industry. These and other issues add complexity and uncertainty to the already
challenging domains of orebody modelling and strategic mine planning, which are arguably the
backbone of our industry.
It is a privilege to introduce the new Spectrum Series Volume 17: Advances in Orebody Modelling and
Strategic Mine Planning I, focusing on ‘Old and new dimensions in a changing world’. This
publication presents the latest advances, technologies, practices and concerns in strategic mine planning
and orebody modelling, thus updating the developments and applications reported in the remarkably
successful 2005 Spectrum Series 14 and its second revised edition in 2007.
The unique qualities of this volume are as evident today as they were in the previous Spectrum Series
volume, thanks to many factors. To begin with, this volume has an impressive author list of
international experts, including a new and promising generation of young colleagues in our field who
we hope to see continuing our efforts in the decades to come; these contributions by so many respected
colleagues is reassuring for our future success. Second, the ongoing support and involvement of the
global mining industry has become indispensable; and finally, the exceptional and continued
collaboration of the major national institutes who have supported this volume and the related
international symposium, held 16 - 18 March 2009, in Perth, Western Australia: The AusIMM
(Australia), CIM (Canada), SME (USA) and The SAIMM (South Africa). In the present global mining
environment, I believe this is a substantial effort by these institutes to enhance professional excellence
in the critical field addressed by this volume. I am proud that this collaboration has been fostered by all
involved, to the benefit of our profession.
This volume comprises 41 papers under the following topics:

• Modelling and Planning in the New World
The volume opens with papers examining global optimisation, mining optimisation and management
of multiple objectives, risk quantification, joint effects of metal and market uncertainties in mine
design, lessons learned and old and new paradigms.
• Modelling Orebodies and Applications
This section stresses new developments for modelling orebodies in different ways, including complex
geology, indirect data and optimising drilling, while keeping in mind the needs of the industry for
methods and implementations facilitating their routine use in the industry environment.
• Mining Optimisation and Case Studies
This group of papers complements the previous section, stressing case studies involving different
commodities.
• Large-Scale Integrated Applications
This section presents major case studies that use new technologies and new approaches, along with
their applications, which include open pit mine design, production scheduling under uncertainty,
cut-off grade optimisation, modelling the supply chain and assessing deep sited gold reserv es.
• Uncertainty Models and Recoverable Reserves
This section is focused on the related issues underpinning the sustainability of the mining industry
and includes new, state-of-the-art simulation approaches and case studies.
• Optimisation is Moving Underground
This section reflects a rapid growth in the application of optimisation approaches in various aspects
of underground mining, from access to ventilation as well as production scheduling.
• The New World and Strategic Innovation
This final section deals with the changing world. It stresses major applications and frameworks for
stochastic mine planning optimisation, and extends to optimising under the joint uncertainty of metal,
markets and mining, as well as newest state-of-the-art high-order simulation approaches and
application of simulation.
As always, a volume such as this is due to the combined efforts of several colleagues over several years.
I would particularly like to thank our colleagues and international experts: Jeff Whittle, Gavin Yeates,
Wynand Kleingeld, Jean-Michel Rendu, Rick Allan, Ian Douglas, Vaughan Chamberlain, Edson
Ribeiro, Brian Baird, David Whittle, Malcolm Thurston, Olivier Tavchandjian, Mike O’Brien,
Jean-Yves Cloutier, Peter Ravenscroft, Jaimie Donovan, Peter Stone, Malcolm Thurston, Andre
Journel, Christophe Pierre and our colleagues involved with the COSMO Lab. Their support over the
past several years has been invaluable. I would further like to acknowledge the diligent work of the
reviewers of the papers in this volume who are listed on page iii, and thank the authors for the high
quality of their contributions.
Once again, support from the industry has been remarkable, with an impressive list of major sponsors
whose long-term multifaceted support I gratefully acknowledge: AngloGold Ashanti, Barrick, BHP
Billiton, De Beers, GEMCOM Australia, Newmont and Vale.
I believe the underlying reason for the unprecedented support for this and previous volumes on strategic
mine planning is the critical importance to the mining industry of new leading-edge technologies that
quantify uncertainty and manage risk in decision-making on a global scale. This uncertainty stems from
both the uncertain supply of metal and an uncertain future market demand. Given the high risk/high
reward nature of mining and the current global economic downturn, this volume of proceedings (i)
contributes new technologies that add substantial value to mining assets, and (ii) assists with the related
knowledge dissemination in a rapidly changing world.
It takes an enormous effort to produce a Spectrum Volume like the one in your hands. I would like to
thank The AusIMM and its staff for its continuing support for our area of work, including Miriam Way,
who has been the source of encouragement and leadership behind several professional development
activities over the past decade. Thanks are also in order to Katy Andrews, Kylie McShane, Kristy
Pocock and Alison McKenzie for their tremendous effort (under conditions of editorial uncertainty)
towards this Spectrum Volume and its quality.
On a personal note, this year marks a decade of collaboration with The AusIMM; I would like to
express my gratitude to this remarkable organisation for its contribution to the mining industry and our
profession. Both Australia and the global mining community are lucky to have such an Institute. Thank
you AusIMM!
Roussos Dimitrakopoulos
COSMO – Stochastic Mine Planning Laboratory
McGill University, Montreal
An initiative of
COSMO
Stochastic Mine Planning Laboratory
Sponsors
Contents
Modelling and Planning in the New World
The Global Optimiser Works – What Next? J Whittle 3
Optimal Mining Principles B King 7
Stochastic Mine Planning – Methods, Examples and Value in an R Dimitrakopoulos 13

Uncertain World
A Risk Analysis Based Framework for Strategic Mine Planning M Godoy 21

and Design – Method and Application
Accounting for Joint Ore Supply, Metal Price and Exchange Rate S A Abdel Sabour and 29
Uncertainties in Mine Design R Dimitrakopoulos
Production Scheduling Under Metal Uncertainty – Application of A Leite and R Dimitrakopoulos 35

Stochastic Mathematical Programming at an Open Pit Copper
Mine and Comparison to Conventional Scheduling
Modelling Orebodies and Applications
Regional Three-Dimensional Modelling of Iron Ore Exploration V Osterholt, O Herod and 43

Targets H Arvidson
Conditional Simulation of Grades Controlled by Geological C Xu and P A Dowd 51

Indicators
Mining Geostatistics with the Open Source SGeMS Software and A Boucher 59
the GsTL Library
Conditional Simulation Methods to Determine Optimum Drill Hole C Boyle 67

Spacing
Transformation Methods for Multivariate Geostatistical Simulation E M Bandarian, U A Mueller, 79

– Minimum/Maximum Autocorrelation Factors and Alternating J Fereira and S Richardson
Columns Diagonal Centres
Conditional Wavelet Based Simulation of Non-Stationary E Gloaguen and 91

Geologic Fields Using Geophysical and Model Analogue R Dimitrakopoulos
Information
Mining Optimisation and Case Studies
Optimal Life-of-Mine Scheduling for a Bauxite Mine M Zuckerberg, J van der Riet, 101
W Malajczuk and P Stone
Coal Mining in the Hunter Valley – A Strategic Mine Planning C Wharton 107
Case Study
A New Methodology for Flexible Mine Design B Groeneveld, E Topal and 113
B Leenders
Pit Optimisation and Mine Scheduling at the Montes Claros De B Y Tomaselli, E A Silva, 123
Goiás Nickel Laterite Deposit, Brazil G J de Carvalho Pereira,
J A Hilário, M C Botelho and
P C Rodriguez
Technogenic Iron Ore Formations in Siberia as a Secondary P A Filippov, S A Neverov and 129
Resource for Production of Ferrous, Rare, Non-Ferrous and A A Neverov
Noble Metals
Advancing Strategic Mine Planning – A Software Perspective C Tonkin 135
Large-Scale Integrated Applications
Integrated Strategic Planning at EKATI Diamond Mine B Coutts 143
Stochastic Long-Term Production Scheduling of Iron Ore J Benndorf and 151

Deposits – Integrating Joint Multi-Element Geological Uncertainty R Dimitrakopoulos
Overall Optimisation of Yerilla Project T Dincer and K Reynolds 159
Investigating the Variability of Key Coal Quality Parameters in J Benndorf 167

Continuous Mining Operations when Using Stockpiles
Cut-Off Grade Estimation – Old Principles Revisited – Application J-M Rendu 173
to Optimisation of Net Present Value and Internal Rate of Return
Using Grade Uncertainty to Quantify Risk in the Ultimate Pit S Robins 179
Design for the Sadiola Deep Sulfide Prefeasibility Project, Mali,
West Africa
Uncertainty Models and Recoverable Reserves
Angle Reverse Circulation Versus Blastholes in Ore Control – I Douglas, B Perry, W Hardtke 193
Sampling Quality Versus Data Density – A Case Study and L Allen
Impact of Bench Height, Cut-Off Grade and Production Spacing W Hardtke, L Allen, B Perry 199
on Misclassification in Ore Control – Challenges Produced by and
High Nugget, Low Continuity Ores I Douglas
A Geology-Based Joint Conditional Simulation of a Lateritic J A Lopes, C F Rosas, 205

Nickel Deposit – Quantification of Risk in Grade Tonnage Curves J B Fernandes and G A
and Resource Categorisation Vanzela
Resource/Reserve Classification with Integrated Geometric and R Dimitrakopoulos, M Godoy 215

Local Grade Variability Measures and C L Chou
Reducing the Geological Risk in Mining an Orebody by Using C Kemp, P Du Pisani, A Bray 223
Borehole Radar for Strategic Mine Planning and G Chitiyo
Optimisation is Moving Underground
Strategic Optimisation of a Vertical Hoisting Shaft in the Callie M G Volz, M Brazil and 235
Underground Mine D A Thomas
Optimising Value for an Underground Project Configuration T Elkington, R Durham and 241
P Myers
Classification of Mining Methods for Deep Orebodies V Oparin, A Tapsiev and 249
A Freidin
Resource Risk Quantification for a Deep Level Gold Mining T Flitton and R Peattie 257
Operation
An Application of Genetic Algorithms to the Main Ventilation E Acuña, S Hardcastle, L Fava, 263
System Optimisation for Multi-Level Mine Operations – The P Dunn and S Hall
Impact of the Fitness Function
The New World and Strategic Innovation
Applying Sequence Optimisation Technology to Longwall Coal L Rocchi, P Carter and P Stone 271
Mining
Combining Optimisation and Simulation to Model a Supply Chain P Bodon, C Fricke, T 277
from Pit to Port Sandeman and C Stanford
Direct Net Present Value Open Pit Optimisation with Probabilistic A Richmond 285
Models
Pushback Design of Open Pit Mines Under Geological and C Meagher, S A Abdel Sabour 291
Market Uncertainties and R Dimitrakopoulos
Faulted Geological Model Simulation of the Resolution Porphyry G Verly, K Brisebois, W Hart 299
Copper Deposit and J Hammitt
Geologically Enhanced Simulation of Complex Mineral Deposits H Mustapha and 309

Through High-Order Spatial Cumulants R Dimitrakopoulos
Stochastic Mine Planning – Example and Value from Integrating A Jewbali and R 321
Long- and Short-Term Mine Planning Through Simulated Grade Dimitrakopoulos
Control, Sunrise Dam, Western Australia
Author Index 329

MONOGRAPH SERIES
1 • Detrital Heavy Minerals in Natural Accumulates George Baker 1962
2. • Research in Chemical and Extraction Metallurgy Ed: J T Woodcock, 1967
A E Jenkins
and G M Willis
3. • Broken Hill Mines - 1968 Ed: M Kadmanovich 1968
and J T Woodcock
4. • Economic Geology of New Zealand Ed: G J Williams 1974
5. • Economic Geology of Australia and Papua New Guinea - 1 Metals Ed: C L Knight 1975
6 • Economic Geology of Australia and Papua New Guinea - 2 Coal Ed: D M Traves 1975
and D King
7. • Economic Geology of Australia and Papua New Guinea - 3 Petroleum Ed: R B Leslie 1976
H J Evans
and C L Knight
8. • Economic Geology of Australia and Papua New Guinea - 4 Industrial Ed: C L Knight 1976
Minerals and Rocks
9. Field Geologists’ Manual
1st Edition Ed: D A Berkman 1976
2nd Edition and W Ryall 1982
3rd Edition Ed: D A Berkman 1989
4th Edition Ed: D A Berkman 2001
10 • Mining and Metallurgical Practices in Australasia Ed: J T Woodcock 1980
(the Sir Maurice Mawby Memorial Volume)
11. • Victoria’s Brown Coal - A Huge Fortune in Chancery Ed: J T Woodcock 1984
(the Sir Willis Connolly Memorial Volume)
12. • Australasian Coal Mining Practice
1st Edition Ed: C H Martin 1986
2nd Edition Ed: C H Martin 1993
and A J Hargraves
3rd Edition Ed: R J Kininmonth 2009
and E Y Baafi
13. • Mineral Deposits of New Zealand Ed: Dr D Kear 1989
14. Geology of the Mineral Deposits of Australia and Papua New Guinea Ed: F E Hughes 1990
15. The Rocks Speak H King 1989
16. • Hidden Gold - The Central Norseman Story J D Campbell 1990
17. • Geological Aspects of the Discovery of Some Important K R Glasson 1990
Mineral Deposits in Australia and J H Rattigan
18. • Down Under - Mineral Heritage in Australasia Sir Arvi Parbo 1992
19. Australasian Mining and Metallurgy Ed: J T Woodcock 1993
(the Sir Maurice Mawby Memorial Volume) and K Hamilton
20. • Cost Estimation Handbook for the Australian Mining Industry Ed: M Noakes 1993
and T Lanz
21. • History of Coal Mining in Australia Ed: A J Hargraves, 1993
(the Con Martin Memorial Volume) R J Kininmonth,
C H Martin
and S M C Saywell
22. Geology of Australian and Papua New Guinean Mineral Deposits Ed: D Berkman 1998
and D Mackenzie
Copies of all books currently in print can be obtained from
The Institute office - Tel (03) 9662 3166
Key: • Out of print

23. Mineral Resource and Ore Reserve Estimation — Ed: A C Edwards 2001
The AusIMM Guide to Good Practice
24. Australian Mineral Economics Ed: P Maxwell and P Guj 2006
25. Geology and Exploration of New Zealand Mineral Deposits Ed: A B Christie and 2006
R L Brathwaite
Copies of all books currently in print can be obtained from
The Institute office - Tel (03) 9662 3166
Key: • Out of print

PUBLICATIONS OF THE AUSTRALASIAN INSTITUTE
OF MINING AND METALLURGY
CONFERENCE, SYMPOSIUM AND MISCELLANEOUS

1972 1/72 SI * Project Evaluation and Management, Melbourne
2/72 * Annual Conference, Newcastle
1973 1/73 S2 * Subsidence, Illawarra
2/73 S3 * Mine Filling, North West Queensland
3/73 S4 * Transportation, Sydney
4/73 S8 * Mine Fires, Southern Queensland
5/73 * Annual Conference, Western Australia
1974 1/74 S5 * Support in Pillar Extraction, Illawarra
2/74 S6 * Recent Technical and Social Advances in the North Australian Minerals Industry,
North West Queensland
3/74 S7 * Pellets and Granules, Newcastle and District
4/74 * Annual Conference, Southern and Central Queensland
1975 1/75 S9 * People and the Mining Industry - The Future, Broken Hill
2/75 S 10 * Occupational Safety in Mines, Southern Queensland
3/75 S 11 * Australian Black Coal, lllawarra
4/75 * Annual Conference, South Australia
1976 1/76 S13 * Landscaping and Land Use Planning as Related to Mining Operations, Adelaide
2/76 * Design and Construction of Tunnels and Shafts, Melbourne - 2 nd Australian
Tunnelling Conference
3/76 S 14 * Thick Seam Mining by Underground Methods, Central Queensland
4/76 S 15 * Sampling Practices in the Mineral Industries, Melbourne
5/76 * Annual Conference, lllawara
1977 1 /77 S 16 * Apcom ‘77, Brisbane
2/77 S 18 * Coal Borehole Evaluation, Southern Queensland
3/77 S 17 * Underground Operators’ Conference, Broken Hill
4/77 * Annual Conference, Hobart
1978 1 /78 S 19 * Mill Operators’ Conference, North West Queensland
2/78 S20 * Rock Breaking Equipment and Techniques, Melbourne
3/78 * International Resource Management, Canberra
4/78 * Annual Conference, Townsville
1979 1/79 S21 * Utilisation of Steelplant Slags, lllawarra
2/79 S22 * Estimation and Statement of Mineral Reserves, Sydney
3/79 * Annual Conference, Perth
1980 1/80 * Annual Conference, New Zealand
2/80 S23 * Australia/Japan Extractive Metallurgy Symposium, Sydney
3/80 S/24 * Occurrence, Prediction and Control of Outbursts in Coal Mines, Southern
Queensland
4/80 S25 * Management in the Mining Industry, Melbourne
Copies of all books currently in print can be obtained from The Institute office
Tel (03) 9662 3166 or Fax (03) 9662 3662
* = Out of print
The ‘S’ numbers in the third column refer to an older identifying number for Symposia, the numbers preceeding the ‘S’ number signify the new
publication ordering number.
1981 I/81 S26 * International Blast Furnace Hearth and Raceway Symposium, Newcastle
2/81 * Fourth Australian Tunnelling Conference
3/81 S27 * Ignitions, Explosions and Fires, Wollongong
4/81 * Annual Conference, Sydney
5/81 S28 * Strip Mining 45 Metres and Beyond, Central Queensland
1982 1/82 S29 * Off Highway Truck Haulage Conference, Newman
2/82 S30 * Mill Operators’ Conference, North West Queensland
3/82 S31 * Underground Operators’ Conference, West Coast Tasmania
4/82 * Annual Conference, Melbourne
5/82 S32 * Carbon-ln-Pulp Technology for the Extraction of Gold, Perth and Kalgoorlie,
(Reprinted 1988)
6/82 S33 * Seam Gas Drainage with Particular Reference to the Working Seam, Wollongong
1983 1/83 S34 * Computers in Mining, Southern Queensland
2/83 * Annual Conference, Broken Hill
3/83 S35 * Project Development Symposium, Sydney
4/83 S37 * Ventilation of Coal Mines, Wollongong
5/83 S40 * Principles of Mineral Flotation (The Wark Symposium), Adelaide
1984 1/84 S36 * Metallurgy Symposium, Melbourne
2/84 S38 * Coal and Mineral Sizing, Wollongong
3/84 * Annual Conference, Darwin
4/84 S39 Gold Mining, Metallurgy and Geology, Kalgoorlie
1985 1/85 S41 * Smelting and Refining Operators’ Symposium, North Queensland
2/85 S42 * Underground Operators’ Conference, Kalgoorlie
3/85 * Annual Conference, Brisbane
4/85 S43 * Scientific and Technological Developments in Extractive Metallurgy (G K
Williams Memorial Volume), Melbourne
1986 1/86 * l3th Congress The Council of Mining and Metallurgical Institutions, Singapore, 6
Volumes
2/86 S44 * Selective, Open Pit Gold Mining Seminar, Perth
3/86 S45 * Ground Movement and Control Related to Coal Mining, Wollongong
4/86 S46 * Australia: A World Source of Illmenite, Rutile, Monazite and Zircon Conference,
Perth
5/86 S47 * Second Project Development Symposium, Sydney
6/86 S48 * Large Open Pit Mining Conference, Newman
7/86 S49 * Education and Research for the Mineral Industry for the Future, Melbourne
8/86 * The AuslMM 10 Year lndex
1987 1/87 * Vl Australian Tunnelling Conference: Bore or Blast, Melbourne
2/87 S50 * Risk and Survival Seminar, Canberra
3/87 * Annual Conference, Newcastle: Coal Power ‘87
4/87 S51 * Research and Development in Extractive Metallurgy, Adelaide
5/87 * Leslie Bradford Golden Jubilee Oration
6/87 S52 * Mining and Environment: A Professional Approach, Brisbane
7/87 * Pacrim ‘87, Gold Coast, Queensland
8/87 S53 * Dense Medium Operators’ Conference, Brisbane
Tel (03) 9662 3166 or Fax (03) 9662 3662
* = Out of print
9/87 S54 * Equipment in the Minerals Industry: Exploration Mining and Processing
Conference, Kalgoorlie
10/87 S55 * Resources and Reserves, Sydney
11/87 * South Australia’s Mining Heritage
1988 1/88 S56 * 21st Century Higher Production Coal Mining Systems Symposium, Wollongong
2/88 * The Second International Conference on Prospecting in Arid Terrain, Perth
3/88 S57 * Third Mill Operators’ Conference, Cobar
4/88 S58 * Underground Operators’ Conference, Mount Isa
5/88 * Fourth International Mine Ventilation Congress, Brisbane (Proceedings and
Addendum volume)
6/88 * Annual Conference, Sydney: Minerals and Exploration at the Crossroads: The
International Outreach
7/88 S59 * Second AuslMM Mineral Heritage Seminar, Sydney
8/88 S60 * Economics and Practice of Heap Leaching in Gold Mining Workshop, Cairns
9/88 * Third International Mine Water Congress, Melbourne
10/88 S61 * Explosives in Mining Workshop, Melbourne
1989 1/89 * Mineralogy and Petrology, Sydney, February
2/89 * Second Large Open Pit Mining Conference, Latrobe Valley Vic
3/89 * NQ Gold ‘89 Conference, Townsville Qld
4/89 * Annual Conference, Perth-Kalgoorlie: Education, Training and Professional
Development; Industrial Minerals; Project Development/Processing
5/89 * Mineral Fuel Alternatives and the Greenhouse Effect, July 1989
6/89 Non-ferrous Smelting Symposium: 100 Years of Smelting and Refining
Operations in Port Pirie, SA September 1989
7/89 * Dewatering Technology and Practice Conference, Brisbane October 1989
8/89 MINVAL ‘89, Mining and Petroleum Valuation 1989, Sydney September 1989
1990 1/90 * Ore Reserve Estimates - The Impact on Miners and Financiers, Melbourne, March
1990
2/90 Annual Conference, The Mineral Industry in New Zealand, Rotorua New
Zealand, March 1990
3/90 Pacific Rim Congress, Gold Coast Qld, May 1990
4/90 * Mining Industry Capital and Operating Cost Estimation Conference, Sydney,
June 1990
5/90 * Third International Symposium on Rock Fragmentation by Blasting, Brisbane,
August 1990
6/90 * Sir Edgeworth David Memorial Oration, May 1990
7/90 * Mine Geologists’ Conference, Mount Isa, October 1990
1991 1/91 * Fourth Mill Operators’ Conference, Burnie Tas, March 1991
2/91 * World Gold ‘91, Cairns Qld, April 1991
3/91 Mining Industry Optimisation Conference, Sydney, June l991
4/91 * PNG Geology, Exploration and Mining Conference, Rabaul, June 1991
5/91 Qld Coal Symposium, Brisbane, August 1991
6/91 * Reliability Production and Control in Coal Mines, Wollongong, September 1991
7/91 * Fifth AuslMM Extractive Metallurgy Conference, Perth, October 1991
Tel (03) 9662 3166 or Fax (03) 9662 3662
* = Out of print
1992 1/92 * Enviromine Australia, Sydney NSW, March 1992
2/92 The AuslMM Annual Conference, ‘The State-of-the-Art - A Product of 100 Years
of Learning’, Broken Hill NSW, May 1992
3/92 ‘Energy, Economics and Environment’ Gippsland Basin Symposium, Melbourne,
June 1992
4/92 Arnold Black Mineral Heritage Oration
5/92 The Man from ASARCO: a life and times of Julius Kruttschnitt
6/92 5th Underground Operators’ Conference, Ballarat, July 1992
7/92 * 11th International Conference on Ground Control in Mining, Wollongong, July
1992
8/92 * Third Large Open Pit Mining Conference, Mackay, August 1992
9/92 * Extractive Metallurgy of Gold and Base Metals Conference, Kalgoorlie, October
1992
10/92 * Sampling Practices in the Minerals Industry, Mount Isa, November 1992
11/92 * Rehabilitate Victoria, Latrobe Valley, November 1992
1993 1/93 Mining People - A Century
2/93 The AuslMM Centenary Conference, Adelaide, March 1993
3/93 XVIII International Mineral Processing Congress, Sydney, May 1993
4/93 * Narrow Vein Mining Seminar, Bendigo, June 1993
5/93 * International Mining Geology Conference, Kalgoorlie, July 1993
6/93 Vlll Australian Tunnelling Conference, Sydney, August 1993
7/93 * World Zinc ‘93 - International Symposium, Hobart, October 1993
1994 1/94 1994 AuslMM Student Conference, Brisbane, April 1994
2/94 * PNG Geology, Exploration and Mining Conference, Lae, PNG, June 1994
3/94 No Two The Same by Bert Mason
4/94 Sixth Extractive Metallurgy Conference, Brisbane July 1994
5/94 * 1994 AuslMM Annual Conference, Darwin, August 1994
6/94 * 4th Large Open Pit Mining Conference, Perth, September 1994
7/94 * Recent Trends in Heap Leaching, Bendigo, September 1994
8/94 * Maintenance in the Mining and Metallurgical Industries,Wollongong, October
1994
9/94 * Fifth Mill Operators’ Conference, Roxby Downs, October 1994
10/94 Mineral Valuation Methodologies 1994, Sydney, October 1994
11/94 Victorian Mining Week Conference, Melbourne, October 1994
1995 1/95 1995 AuslMM Annual Conference, Newcastle, March 1995
2/95 Sir Maurice Mawby Memorial Oration
3/95 World’s Best Practice in Mining and Processing Conference, Sydney, May 1995
4/95 APCOM XXV 1995 Conference, Brisbane, July 1995
5/95 Mineral Valuation Methodologies 1994, Sydney, October 1994 (revised)
6/95 * EXPLO 95 Conference, Brisbane, September 1995
7/95 Underground Operators’ Conference, Kalgoorlie, November 1995
8/95 * Young Professionals’ Conference, Mt Isa, October 1995
9/95 * PACRIM ‘95 Congress, Auckland, New Zealand, November 1995
10/95 * Ethics, Liability and the Technical Expert, Sydney, December 1995
Tel (03) 9662 3166 or Fax (03) 9662 3662
* = Out of print
1996 1/96 * 1996 AusIMM Annual Conference, Perth, March 1996
1a/96 * 1996 AusIMM Annual Conference Supplementary Volume, Perth, March 1996
2/96 * Ethics, Liability and the Technical Expert, Sydney, March 1996
3/96 * Entrepreneurs and Partners, Sydney, July 1996
4/96 * Contract Operators’ Conference, Kalgoorlie, October 1996
5/96 Asia/Pacific Mining Communications Summit, Singapore, November 1996 -
Withdrawn
6/96 * Nickel ‘96, Kalgoorlie, November 1996
1997 1/97 1997 AusIMM Annual Conference, Ballarat, March 1997
2/97 * World Gold ‘97 Conference, Singapore, September 1997
3/97 Sixth Mill Operators’ Conference, Madang, PNG, October 1997
4/97 Gem 97, Madang, PNG, October 1997
5/97 * Contract Operators’ Conference, Brisbane, Qld, October 1997
6/97 Third International Mining Geology Conference, Launceston, Tas, November
1997
7/97 Mindev 97 - The International Conference on Mine Project Development,
Sydney, November 1997
8/97 1997 AusIMM Travelling Technology Forum, Singleton, NSW, March 1997
1998 1/98 * MINEFILL ‘98 - The Sixth International Symposium on Mining with Backfill,
Brisbane, Qld, April 1998
2/98 * AusIMM’98 - The Mining Cycle, Mount Isa, Qld, April 1998
3/98 * Seventh Underground Operators’ Conference, Townsville, Qld, June/July 1998
4/98 Mine to Mill Conference, Brisbane, Qld, October 1998
5/98 Third Regional APCOM - Computer Applications in the Minerals Industries
International Symposium, Kalgoorlie, WA, December 1998
6/98 COAL98 - First Australasian Coal Operators’ Conference, Wollongong, NSW,
February 1998
1999 1/99 10th Australian Tunnelling Conference, Melbourne, Vic, March 1999
1a/99 10th Australian Tunnelling Conference Keynote Addresses and Asia–Pacific
Forum, Melbourne, Vic, March 1999
2/99 Students and Young Professionals Conference, Perth, WA, July 1999 (N/A)
3/99 ICARISM ’99 Conference, Perth, WA, September 1999
4/99 PACRIM ’99 Congress, Bali, October 1999
5/99 EXPLO ’99 Conference, Kalgoorlie, WA, November 1999
2000 1/2000 Southern Africa - Australia Mineral Sector Synergies Symposium, Canberra,
ACT, March 2000
2/2000 After 2000 - The Future of Mining, Sydney, NSW, April 2000
3/2000 4th International Mining Geology Conference, Coolum, Qld, May 2000
4/2000 Young Leaders 2000, Sydney, NSW, July 2000
5/2000 MINPREX 2000, Melbourne, Vic, September 2000
6/2000 Seventh Mill Operators’ Conference, Kalgoorlie, WA, October 2000
7/2000 MassMin 2000, Brisbane, Qld, October - November 2000
2001 1/2001 Strategic Mine Planning 2001, Perth, WA, March 2001
2/2001 AusIMM Youth Congress 2001, Brisbane, Qld, May 2001
3/2001 International Heavy Minerals Conference, Fremantle, WA, June 2001
4/2001 EXPLO 2001, Hunter Valley, NSW, October 2001
Tel (03) 9662 3166 or Fax (03) 9662 3662
* = Out of print
5/2001 Mineral Asset Valuation Issues 2001, Sydney, NSW, October 2001
6/2001 Coal2001, Coal Mining Operators’ Geotechnology Colloquium, Wollongong,
NSW, February 2001
2002 1/2002 Young Leaders 2002, Kalgoorlie, WA, March 2002
2/2002 Metallurgical Plant Design and Operating Strategies, Sydney, NSW, April 2002
3/2002 CMMI Congress, Cairns, Qld, May 2002
4/2002 Green Processing Conference, Cairns, Qld, May 2002
5/2002 Eighth Underground Operators’ Conference, Townsville, Qld, July 2002
6/2002 AusIMM 2002 Conference, 150 Years of Mining, Auckland, New Zealand,
September 2002
7/2002 Iron Ore Conference, Perth, WA, September 2002
8/2002 Value Tracking Symposium, Brisbane, Qld, October 2002
9/2002 Coal2002, Third Australasian Coal Operators’ Conference, Wollongong NSW,
February 2002
2003 1/2003 Twelfth International Symposium on Mine Planning and Equipment Selection,
Kalgoorlie, WA, April 2003
2/2003 Young Leaders 2003, Brisbane, Qld, April 2003
3/2003 Sixth International Conference Acid Rock Drainage, Cairns, Qld, July 2003
4/2003 Eighth Mill Operators’ Conference, Townsville, Qld, July 2003
5/2003 Mining Risk Management Conference, Sydney, NSW, September 2003
6/2003 Water in Mining 2003, Brisbane, Qld, October 2003
7/2003 Fifth Large Open Pit Conference, Kalgoorlie, WA, November 2003
8/2003 Fifth International Mining Geology Conference, Bendigo, Vic, November 2003
9/2003 Coal2003, Fourth Australasian Coal Operators’ Conference, Wollongong NSW,
February 2003
2004 1/2004 New Leaders 2004, Ballarat, Vic, April 2004
2/2004 Second International Green Processing Conference, Fremantle, WA, May 2004
3/2004 EXPLO 2004, Perth, WA, July 2004
4/2004 Metallurgical Plant Design and Operating Strategies, Perth, WA, September 2004
5/2004 PACRIM 2004 Congress, Adelaide, SA, September 2004
6/2004 Bac-Min Conference, Bendigo, Vic, November 2004
7/2004 Orebody Modelling and Strategic Mine Planning, Perth, WA, November 2004
8/2004 Coal2004, Fifth Australasian Coal Operators’ Conference, Wollongong, NSW,
February 2004
2005 1/2005 Ninth Underground Operators’ Conference, Perth, WA, March 2005
2/2005 Coal2005, Sixth Australasian Coal Operators’ Conference, Brisbane, Qld, April
2005
3/2005 New Leaders 2005, Brisbane, Qld, April 2005
4/2005 Second World Conference on Sampling and Blending, Sunshine Coast, Qld, May
2005
5/2005 Centenary of Flotation Symposium, Brisbane, Qld, June 2005
6/2005 Eighth International Mine Ventilation Congress, Brisbane, Qld, July 2005
7/2005 Hoist and Haul Conference, Perth, WA, September 2005
8/2005 Iron Ore Conference, Fremantle, WA, September 2005
9/2005 First Extractive Metallurgy Operators’ Conference, Brisbane, Qld, November 2005
Tel (03) 9662 3166 or Fax (03) 9662 3662
* = Out of print
2006 1/2006 Disposal of Mining Waste Symposium, Sydney, NSW, March 2006
2/2006 New Leaders 2006, Kalgoorlie, WA, April 2006
3/2006 Third International Green Processing Conference, Newcastle, NSW, June 2006
4/2006 Coal2006, Seventh Underground Coal Operators’ Conference, Wollongong,
NSW, July 2006
5/2006 Australia’s Uranium, Adelaide, SA, July 2006
6/2006 Sixth International Mining Geology Conference, Darwin, NT, August 2006
7/2006 AusIMM New Zealand Branch 2006 Conference, Waihi, New Zealand, August
2006
8/2006 Metallurgical Plant Design and Operating Strategies, Perth, WA, September 2006
9/2006 International Mine Management Conference, Melbourne, Vic, October 2006
10/2006 Water in Mining 2006, Brisbane, Qld, November 2006
2007 1/2007 Ninth Mill Operators’ Conference, Fremantle, WA, March 2007
2/2007 New Leaders 2007, Brisbane, Qld, May 2007
3/2007 Australia’s Uranium, Darwin, NT, May 2007
4/2007 Project Evaluation Conference, Melbourne, Vic, June 2007
5/2007 AusIMM New Zealand Branch 2007 Conference, Christchurch, New Zealand,
August 2007
6/2007 Iron Ore Conference, Perth, WA, August 2007
7/2007 EXPLO 2007, Wollongong, NSW, September 2007
8/2007 6th Large Open Pit Mining Conference, Perth, WA, September 2007
9/2007 World Gold 2007, Cairns, Qld, October 2007
10/2007 Uranium Reporting Workshop, Adelaide, SA, November 2007
2008 1/2008 Coal2008, Eighth Underground Coal Operators’ Conference, Wollongong, NSW,
February 2008
2/2008 Tenth Underground Operators’ Conference, Launceston, Tas, April 2008
3/2008 13th Australian Tunnelling Conference, Melbourne, Vic, May 2008
4/2008 Sampling 2008 Conference, Perth, WA, May 2008
5/2008 AusIMM International Uranium Conference, Adelaide, SA, June 2008
6/2008 The AusIMM New Leaders’ Conference 2008, Wollongong, NSW, July 2008
7/2008 MetPlant 2008, Perth, WA, August 2008
8/2008 Ninth International Congress for Applied Mineralogy (ICAM) 2008, Brisbane,
Qld, September 2008
9/2008 Narrow Vein Mining Conference 2008, Ballarat, Vic, October 2008
10/2008 First International Future Mining Conference and Exhibition 2008, Sydney,
NSW, November 2008
11/2008 PACRIM Congress 2008, Gold Coast, Qld, November 2008
2009 1/2009 Coal2009, Ninth Underground Coal Operators’ Conference, Wollongong, NSW,
February 2009
2/2009 Orebody Modelling and Strategic Mine Planning, Perth, WA, March 2009
3/2009 Project Evaluation 2009 Conference, Melbourne, Vic, April 2009
4/2009 The AusIMM New Leaders’ Conference 2009, Brisbane, Qld, April 2009
5/2009 AusIMM International Uranium Conference, Darwin, NT, June 2009
6/2009 Sustainable Development Indicators in the Minerals Industry (SDIMI) 2009,
Gold Coast, Qld, July 2009
Tel (03) 9662 3166 or Fax (03) 9662 3662
* = Out of print
7/2009 Iron Ore 2009 Conference, Perth, WA, July 2009
8/2009 Seventh International Mining Geology Conference 2009, Perth, WA, August
2009
9/2009 AusIMM New Zealand Branch Conference 2009, Queenstown, New Zealand,
August 2009
10/2009 Water in Mining 2009, Perth, WA, September 2009
11/2009 Tenth Mill Operators’ Conference, Adelaide, SA, October 2009
2010 1/2010 Tenth Underground Coal Operators’ Conference, Wollongong, NSW, February
2010
2/2010 The AusIMM New Leaders’ Conference 2010, Kalgoorlie, WA, April 2010
3/2010 Sampling 2010 Conference, Perth, WA, May 2010
4/2010 AusIMM International Uranium Conference, Adelaide, SA, June 2010
5/2010 Seventh Large Open Pit Mining Conference, Perth, WA, July 2010
Tel (03) 9662 3166 or Fax (03) 9662 3662
* = Out of print
SPECTRUM SERIES
1. Making the Mount Isa Mine, 1923 - 1933 Don Berkman 1996
2. History of Drilling Graham McGoggan 1996
3. The Cobar Mineral Field – A 1996 Perspective Warren Cook 1996
Andrew Ford
Julian McDermott
Peter Standish
Craig Stegman and
Therese Stegman
4. Towards 2000 – Resource to Reserve Inputs Seminar, Melbourne, Vic 1997
5. Towards 2000 – National Conference on Ironmaking Resources and Reserves 1997

Estimation, Perth, WA
6. Towards 2000 – The Resource Database Towards 2000, Wollongong, NSW 1997
7. Towards 2000 – Ore Reserves and Finance, Sydney, NSW 1998
8. Towards 2000 – Assessment of Reserves in Low Rank Coals, Morwell, Vic 1997
9. Towards 2000 – Ore Reserve Reconciliation Workshop, Darwin, NT 1997
10. Towards 2000 – Gold and Nickel Ore Reserve Estimation Practice Seminar 1998
11. Towards 2000 – Resource/Reserves Estimation Practice in the Central West 1999
New South Wales Mining Industry, Cobar, NSW
12. Field Guide for Geoscientists and Technicians
1st Edition 2004
2nd Edition 2007
13. The Extractive Metallurgy of Zinc Roderick J Sinclair 2005
14. Orebody Modelling and Strategic Mine Planning – Uncertainty and Risk Roussos Dimitrakopoulos
Management Models
1st Edition 2005
2nd Edition 2007
15. The Extractive Metallurgy of Lead Roderick J Sinclair 2009
16. Flotation Plant Optimisation – A Metallurgical Guide to Identifying and Ed: Christopher J Greet 2010
Solving Problems in Flotation Plants
17. Advances in Orebody Modelling and Strategic Mine Planning I – Old and Roussos Dimitrakopoulos 2010
New Dimensions in a Changing World
HOME
The Global Optimiser Works – What Next?

J Whittle1
ABSTRACT the life of the complex. Expenses are delayed and revenues are
brought forward, while meeting all the required constraints. In
The Global Optimiser used by Whittle Consulting has gone through three
major versions to date. The first was based on the Milawa optimisation
short-term scheduling, practicality and convenience are more
algorithm; it worked, but had many shortcomings. The second, known important than the timing of cash flows. There is also the
internally as ProberA, had a different approach to optimisation in that it consideration that the data required for short-term scheduling
used a series of random starting points and found the nearest local NPV differs from that for long-term scheduling. For long-term
maximum to each. It did this a sufficient number of times to give us some scheduling what matters is the total volume of dirt a truck can
confidence that the best result found was close to optimal. ProberB was move in a year; however, there is no need to consider its detailed
an enhanced version of ProberA, with the ability to handle a wider range movements.
of constraints, particularly with regard to limits on the differences in For Prober, the resource is modelled as a number of
depth between adjacent areas of a pit. ProberB has been used successfully ‘sequences’ of ‘panels’, and the rule is that mining of a panel
for some time now. It has produced excellent Life of Project schedules for
a wide range of very large mining complexes. However, like any piece of
can’t start until the panel before it in the sequence has been
software, it has its limitations. For example, it only copes directly with completely mined. These panels usually represent benches in a
three steps – mining, processing, and blending. push-back, but can also represent parts of an underground mine
It is possible to ‘fool’ the program into handling other steps, but only that must be accessed in a particular order. A panel consists of
by using mental and mathematical gymnastics. This paper reviews the one or more ‘parcels’ of material, each of which will typically
mechanisms behind the Prober series and describes the plans for the next have a material type, a tonnage and a number of grades. If Prober
version – ProberC. mines only a fraction of a panel in a particular year, it uses the
mathematical fiction that the same fraction of every parcel in the
INTRODUCTION panel will be mined in that year. Of course, this is not what
will happen in the day-to-day scheduling, but the necessary
For some years now a global optimiser has been used to produce adjustments should have little effect on the NPV.
long-term mining and processing schedules which maximise the
Net Present Value (NPV) of the whole project, taking account of The main data input to Prober consists of a ‘resource tree’
all cash flows including capital expenditure. There have been showing all possible things that can happen to material after it
three versions of this optimiser, known internally as Z3, ProberA has been mined. For each parcel, it gives the alternative
and ProberB. Further details can be found in Whittle and Whittle processing paths that can be taken (eg mill or heap leach). For
(2007) and in Whittle (2007). This paper explains the methods each processing path, it gives the products that the outputs from
used in Prober and describes the plans for the development of a processing can be blended into, etc. The remainder of the data
new optimiser – ProberC. details cost and price factors apply to the whole complex,
together with the various operational constraints. When
MODELLING modelling the constraints on the operation of the complex, it is
not necessary to simplify much at all. Limits can be set on
In order to use a computer to optimise any system it is necessary mining and processing throughputs, which can depend not only
to create a mathematical model of the system. This model must on tonnage but on grades or other characteristics (eg how much
represent the system as accurately as possible, but also be power or acid is required to process material). Upper and lower
amenable to optimisation. In theory any mathematical model can limits can be set for the average grades input to processing and in
be optimised by throwing enough computer power at it. products, as well as on the depth separation between sequences.
However, as the size of models increases, some types can get
rapidly out of hand. There are mathematical problems which are
quite easy to describe but could not be fully optimised by using THE PROBLEM OF LOCAL MAXIMA
all the computers in the world for a million years. However, it is
sometimes possible to find very good near-optimal solutions to The aim is to find the mining and processing schedule which
problems that defy full optimisation. Occasionally this can be gives the highest NPV. Consider just one sequence and one year
done by simplifying the model, other times by using approximate with no blending. This might result in the situation illustrated in
optimisation methods. The Prober approach uses both these Figure 1. Sequence A has some panels, some of which are ore
techniques to obtain good, long-term schedules for large mining and others are waste. As mining proceeds through the ore, the
complexes. Keep in mind that all models are wrong but some are value goes up; as it proceeds through waste, it goes down. The
useful (Box, 1979). result in this case is that there are two peaks. Figure 2 illustrates
A large mining complex with 20 or more pits with associated the same idea for another sequence. Each peak is a maximum in
processing plants and infrastructure would defy Life of Project its own part of the graph, but it is a local maximum. Only one is
NPV optimisation if modelled in day-to-day detail. It is therefore the global maximum. Figure 3 shows what happens if the
necessary to simplify the model to some degree. The main complexity is increased to two sequences, but still only over one
simplification is to concentrate on long-term scheduling. If a year. Now there are four local maxima. In practice, with a few
large expense is delayed by five years, this could have a dozen pits, blending constraints and many years, there can be
significant effect on the NPV. If the same expense is delayed by a
thousands of local maxima. This can be equated to the search for
week, the effect will be small. Long-term scheduling decides
the highest point in the Himalayas rather than the highest point
what will be mined, processed, blended and sold in each year of
of Mount Fuji. Merely going uphill from a random start point
will get you to a ‘nearest’ local maximum. It is unlikely to be the
1. FAusIMM, Director of Research, Whittle Consulting Pty Ltd, highest point, that is, the global maximum. Because of the
42 Yeneda Street, Balwyn North Vic 3104. multiplicity of local maxima, this is not a linear problem in the
Email: [email protected] sense of linear optimisation – it is a ‘mixed integer’ problem.
Advances in Orebody Modelling and Strategic Mine Planning I Spectrum Series Volume 17 3
J WHITTLE
HOW THE PROBER SERIES OF PROGRAMS

WORK
The Prober series of programs break the problem down into a
number of smaller problems which can each be solved quickly
using linear optimisation software. Prober first produces a
random feasible mining schedule consisting of a year-end depth
in each sequence for each year in the life of the project. Thus, for
about 20 sequences and 15 years, there would be 300 depths.
This schedule is produced by using pseudo-random numbers to
calculate the depths while taking account of all the constraints on
mining throughput and on depth separation. This fixes the
material available for processing in each year. There is no
optimisation in this step and this is not necessarily a good
schedule. Indeed, it is usually very poor, but it is feasible from a
FIG 1 - The total value of material mined from Sequence A plotted mining point of view, so it gives us a starting point. You can
against depth mined, where some benches are ore and some think of this schedule as a random point on the ground
are waste. The two peaks are local maxima; the left peak is the somewhere in the Himalayas. Given the material made available
global maximum. in each year, it is a straightforward matter to use standard linear
optimisation to determine what to process, blend and sell so as to
maximise the NPV obtainable by processing the material subject
to the throughput constraints.
As explained earlier, the assumption is made, so far as the
model is concerned, that parcels within a panel are all mined in
the same proportion. This means that changes in mining depth
within a panel have linear effects on the quantities of material
available for processing and so on. Consequently, in addition to
having the linear optimisation package control what to process,
blend and sell, it is possible to let it control the depths mined,
providing that those depths are constrained to stay within the
panel that they started in. The result is the best mining and
processing schedule for a given set of year-end panels. Having
obtained such a result, the depth details suggest changes to the
year-end panels. For example, if the year-end panel for a
particular sequence and period is set to three, but the
FIG 2 - The total value of material mined from Sequence B plotted optimisation does not mine any of Panel 3, the program will
against depth mined, where some benches are ore and some are change the year-end panel to two and try again. If the NPV
waste. The two peaks are local maxima; the right peak is the improves, the process is repeated. The program is effectively
global maximum. going uphill in the NPV landscape, and this process is repeated
until a maximum is reached. This is the nearest local maximum
to the initial random feasible schedule. One of the peaks in the
Himalayas has been found.
The Prober series of programs works by repeatedly creating a
random feasible solution and then finding the nearest local
maximum. The various NPVs that the program finds are kept track
of, and the run is usually stopped when the top ten values lie
within 0.1 per cent of each other. This does not guarantee
optimality, but it does give confidence that the best results are
pretty close. Most projects that we worked on had a distribution of
local maxima NPVs that is pretty flat near the highest values as
illustrated in Figure 4, which is from a real case. That is, many
different schedules are found that have NPVs which are very
similar to the best. In this case, 11 of the 500 local maxima found
are in the top 0.1 per cent and over 200 are in the top one per cent.
This can be an advantage because it gives managers the
FIG 3 - The total value of material mined when both Sequence A opportunity to choose the schedule that best suits less measurable
and Sequence B are mined to various depths. There are four peaks. considerations such as social, political, environmental and project
Only the rear left peak is the global maximum. This shows just two risk effects. To date, many projects have been successfully
sequences and one year. In real cases with many sequences and completed with the current version of Prober (ProberB) and very
many years, there can be thousands of local maxima. significant NPV gains have been made totalling billions of dollars.
PROBERC
Although linear optimisation methods are very fast for large
problems, the introduction of integers – which is necessary to ProberC uses the same optimisation system but operates on a
deal with the multiple local maxima – slows the process down by greatly enhanced data structure. It can be regarded as a
several orders of magnitude. Indeed, schedules for very large generalisation of ProberB. Grades are expressed as quantities, so
mining complexes are effectively unobtainable using mixed that a grade of 1.2 grams/tonne of gold in a 2000 tonne portion of
integer software. material would be expressed as 2400 grams. Thus, a portion of
4 Spectrum Series Volume 17 Advances in Orebody Modelling and Strategic Mine Planning I
THE GLOBAL OPTIMISER WORKS – WHAT NEXT?
2. a cost or revenue proportional to the difference between a

grade (ratio between two quantities) and a base grade
multiplied by a third quantity; and
3. a constant attached to each portion of material in the
resource tree each time it can be processed or delivered.
All the factors and base grades in components 1 and 2 can vary
independently with the year. Any number of components of types
1 and 2 can apply. Type 3 components are attached to each
portion of material operated on and can be calculated to suit that
portion, but are fixed in time. The scaling factor by which the
sum of these components is multiplied can vary with the year.
This system allows almost any cost or revenue structure to be
modelled. Setting up such costs and revenues might appear
FIG 4 - 500 NPVs obtained by moving from 500 random schedules
somewhat daunting, but it is important to remember that the data
to the nearest local maximum have been sorted into ascending
is always set up by a specialist consultant, not the end user. Costs
order of value. The curve is relatively flat at the high end and 11 of
the local maxima have NPVs that are within 0.1 per cent of the best.
and throughput limits can be applied to both the inputs and the
outputs of procedures. Despite the great increase in complexity
of the processing and cost structures that can be modelled, the
material, be it run-of-mine or the output from processing, is problem steps can each still be handled by a standard linear
described by the quantities associated with it. Any of these optimisation package as in ProberB. ProberC is currently under
quantities can be involved in throughput limits, costs, revenues, development and it is expected to be completed in 2009.
etc. In ProberC, everything that is done to material is done by a
‘procedure’. Mining, processing, smelting, blending, etc are all CONCLUSIONS
procedures. There can be as many procedures as required and
any procedure can provide input to another procedure, with the ProberB has established itself as a powerful and flexible tool.
restriction that material cannot ‘loop’ through the system and ProberC will have the same power as ProberB, but will handle a
thus provide input to the procedure that produced it. Thus, there much wider range of project and costing complexity.
is an arbitrary tree structure of procedures, starting with a mining
procedure and going through as many steps as required to the REFERENCES
point(s) of sale or to waste. This contrasts with ProberB, which
handles only mining, processing and blending for sale. Delivery Box, G E P, 1979. Robustness in the strategy of scientific model building,
in Robustness in Statistics (eds: R L Launer and G N Wilkinson),
of material from a procedure to stockpile, sale, discard or another p 202 (New York: Academic Press).
procedure is much more explicit in ProberC than ProberB and Whittle, G, 2007. Global asset optimisation, in Orebody Modelling and
has its own cost structure. For example, long distance delivery Strategic Mine Planning, second edition (ed: R Dimitrakopoulos),
costs can be handled explicitly rather than having to be included pp 331-336 (The Australasian Institute of Mining and Metallurgy:
in the processing cost as in ProberB. Melbourne).
All costs and revenues are calculated as the sum of one or Whittle, J and Whittle, G, 2007. Global long-term optimisation of very
more of three components, multiplied by a scaling factor. The large mining complexes in Proceedings APCOM 2007 – 33rd
International Symposium on Application of Computers and
components are:
Operations Research in the Mineral Industry, Santiago, pp 253-260.
1. a cost or revenue proportional to one or more of the
quantities, which describe the portion of material involved;
HOME
Optimal Mining Principles

B King1
ABSTRACT There is a subtle distinction between optimising and

maximising. Decisions, policies and strategies are optimised (not
Picture yourself responsible for the exploitation of a world class deposit,
creating staggering quantities of products for global consumption over maximised), Net Present Value (NPV), cash flow or value is
decades, employing thousands of people, about to spend billions of dollars maximised (not optimised). If one were to maximise the cut-off
on infrastructure and you are going to design this project. Should it be grade (to an extreme), then no ore would be sent to the mill and
surface or underground (or both), how big should the processing plants be, the maximum NPV is unlikely to be achieved. On occasion,
what technology should be utilised, what happens if more resources are maximising a policy may also maximise the objective – it is
found or the price forecast changes? This paper aims to help guide better to use the right terminology and sound like an expert!
engineers faced with the prospect of determining optimal mining policies The focus of this paper will be to utilise optimised strategies to
for large projects. It draws on experiences at some of the largest mining
create optimised strategic plans. Given that all strategies discussed
projects and mining companies in the world, including the Bingham
Canyon Mine (USA), Freeport (Irian Jaya), Escondida (Chile), in this paper will refer to optimal mining plans, a Strategic Plan in
Chuquicamata (Chile), Hamersley Iron (Australia), Ekati Diamond Mine this paper is defined as the process of making a number of
(Canada) and several Hunter Valley coal mines (Australia). The paper operating decisions over the project life, to achieve the project
outlines some important areas to put in place before starting, principles to objective. There is clearly a need in the industry to standardise our
guide the planning process and suggestions for finding additional value. terminology and develop an optimal mining practice for the
community. Ideally, this will be a collaborative effort between
INTRODUCTION mining companies and the service companies that support them.
These groups need to span the industry, to help reduce the
misunderstanding that often leads to poor exploitation of the
Terminology earth’s finite resources (King, 2008). The paper demonstrates these
One of the first sources of confusion in strategic planning arises principles using open pit metalliferous projects, although these are
due to the definition of terms (or lack thereof). The terms used in actually generic principles that are appropriately applied to coal,
this paper are based on Ken Lane’s cut-off grade optimisation iron ore, diamonds, bauxite, gold and also many non-mining
work (Lane, 1964, 1988), which was based on Richard Bellman projects.
foundational dynamic programming methodology (Bellman,
1957). Over the last ten years of working with major mining Objectives
projects and feasibility studies, I have found these definitions
useful and applicable in mining and other industries. Mining The first issue that needs to be clarified at the start of a planning
projects have many choices and therefore many decisions to exercise is the work objective. Although there would appear to be
make. A specific decision that is made in every period is grouped a clear and simple answer to the objective, the reality is often
blurred by conflicting instructions. Maximising shareholder
and called a policy. For example, an open pit copper mine may
value is normally the objective of large and small mining
have a decision on mill cut-off grade to be made in every period.
companies (and non-mining companies). This objective is
When the same decision is made at every time period, the result
clarified by the major mining houses as maximising the NPV,
is a constant policy. Typically, different decisions are made over subject to a number of constraints such as safety and good
time, in which case a variable policy is formed. Other operating stewardship of the environment:
policies that may change over time would normally include how
much material is moved from a certain area, SAG mill grind size • NPV is a sound basis for evaluating companies outside of the
and flotation residence time; each needs to be made in every mining world. For example, stock market analysts will often
period. value the share price for a company (in any industry) by
calculating the NPV of all of the company assets, then
Mining studies must typically determine many different
dividing it by the number of shares issued (Brealey and Myers,
policies during the project life. A group of policies is referred to 2000). NPV includes a balanced valuation of short- term value
as a strategy. For example, mining policies for a number of (which receives very little discounting) and long- term value
pushbacks could be described as a mining strategy. By (which receives greater discounting), making it an appropriate
combining a number of mining and processing policies, we can instrument for the commercial valuation of most businesses.
arrive at the strategy for the project, or a strategic plan. A The main difficulties reported with valuing a project based on
decision, policy or strategy does not imply any optimisation NPV include uncertain information and Net Present Value
criteria have been used to define them – it simply defines a versus present value.
number of decisions. Sadly, in real projects, many ‘strategic’
plans also have very little optimisation applied. In order to • With regards to uncertain information, typically there is
‘optimise’ a decision (or policy or strategy), we need to have an substantial uncertainty in key drivers of the NPV including
objective or objective function. The objective will be discussed prices, costs, reserves and productivities. For example, price
is normally a huge driver of value, and financial analysts
later in this paper, but for now, let’s assume it is some definition
recognise that price is related to supply and demand.
of shareholder value. An optimisation process is a way of making
Theoretically, as different strategies are evaluated with
decisions to achieve the project objective. For example, a cut-off
different quantities of metal produced, the prices could
grade optimisation process will optimise the cut-off grade policy change for each case. The complexity does not stop here –
so that the Net Present Value (objective function) is maximised. the price is dependent on the supply from competitors, which
would require an analysis of what they are doing as well. If
1. Managing Director, Strategy Optimisation Systems Pty Ltd, 24 Harris this is not complex enough, the price is also dependent on the
Street, Wellington Point Qld 4160. demand for end products. This means that a comprehensive
Email: [email protected] estimate of price also requires analysis of what competitors
B KING
and customers are planning. A risk/reward analysis would costs, revenue and constraints. The safe operation of a mine or
normally indicate that a comprehensive competitor/customer processing plant will involve additional cost components and
analysis is not valuable. Although specific aspects of risk are constraints on how the activities can be undertaken. Typically
often incorporated using sensitivity studies of targeted issues, the constraints will involve restrictions in productivity rates.
there is a still substantial development required to turn this Good environmental stewardship may involve continual ground
analysis into useful decision making tools for the mining rehabilitation and water treatment costs – it may also require some
industry. Option value, and the ability to use variability to constraints such as tailings dam or waste dump footprints. Some
add value and control risk, is regularly missing from project care is needed to make sure a cost is not mistakenly implemented
valuation analysis. Advances in mathematical formulation as a constraint.
and simulation methodologies can reveal a more complete
Creating ‘unconstrained’ schedules may be of some temporary
understanding of project value. Incorporation of uncertainty
benefit to understand the value drivers for a project, but be wary
and risk needs to become common place in the industry for
of releasing such plans to a wider audience as it may generate
fair valuation between investment alternatives.
unrealistic expectations. For example, when considering the size
The effort required to truly calculate a ‘Net’ Present Value of a process plant expansion, the mining rates may initially be
means that often important components are ignored and a present left unconstrained. This unconstrained case may be used to get
value is used instead. Key aspects such as period costs, expansion the maximum size of the truck and shovel fleets. However, the
capital, sustaining capital and closure costs are not able to be best process plant capacity, when unconstrained, may not be the
incorporated in some optimisation algorithms. Ignoring such best process plant capacity when constrained. The unconstrained
issues or modelling them inadequately means that the results are cases would tend to over-expand the mill capacity since it does
crude, and value destroying decisions can be made. not consider the purchase costs for the mining equipment. These
While both of these difficulties appear to limit the suitability ‘unconstrained’ schedules and their ‘imaginary’ values are not
of using NPV to value a project or different scenarios within a achievable and so would only be used for finding the boundaries
project, NPV still remains at the heart of project valuations and is for analysis.
normally superior to the other financial instruments such as
internal rate of return, payback period and simple cash flows. Plan the planning
Care should be taken when using other measurements to There could well be only one chance to determine the best value
approximate the value of a project. People not aware of NPVs of a project, then acquisition, disposal or operation. Time, staff,
balancing of short- and long-term cash flows may be presented consultants, computing power and budgets are all limited
with some of the following alternatives for the planning resources that need to be used wisely. Much like the optimisation
objectives: process we apply to a project, we need to identify the resources
• product quantity – especially in the budgeting period (eg needed for some analysis and make sure that key people and
gold or copper produced in the first three to five years, support is available during the study. Experience suggests that
very few people have been exposed to this field of the business,
• mineral resources or reserves (especially stated to the stock and so the need for training before the study is necessary, as well
market), as support for when things do not simply fall into place. One of
• smooth production schedules (eg constant material movement the problems with using inexperienced people is that they often do
and constant grades for processing), not know when additional value is possible from a project. The
new planner can make assumptions that are not easily seen in the
• minimising costs or maximising short-term cash flow, and resulting schedules and so may not be picked up by management
• maximising employment. or peer review.
Using the above alternatives for the objective will normally The unrealised value ‘left on the table’ by inexperienced
result in a different plan than the one created by maximising analysis can easily be several percentage points of the NPV, and
NPV and so will have a lower NPV. Take the objective of for large multi-billion dollar projects these are substantial values.
maximising cash flow as an example. Waste stripping incurs It is therefore essential to have the best people working on these
substantial costs that reduce the cash flow. If waste stripping is projects and have them reviewed by strategic planning specialists
suspended, cash flow for the first few years may be improved. (internal and/or consultants). Real world strategic planning is
Eventually, there will be no ore available to process, the mill normally constrained by time. The time spent designing a model,
collecting and validating inputs, building the model, running
capacity will not be utilised, and the cash flows (and NPV)
cases, analysing the cases and then presenting the results, these
rapidly fall. The NPV can be used to balance the early costs of
activities consume precious and finite time resources. Experience
waste stripping with the benefit of having the mill capacity would suggest that gathering and validating inputs often increases
continuously utilised. All aspects of mining and downstream as input data is collected and checked. Using available information
processing that involve costs or revenue are captured directly in such as budget costs might initially appear to rapidly provide
the objective function (NPV). The better our estimates of these optimisation inputs, but they may need to be substantially re-
costs and revenues, the better our estimates of NPV, which will categorised into fixed and variable components for strategic
in turn help bring value to shareholders. planning purposes. A planned work program will have a
contingency for some unforseen data collection issues and changes
Constraints in direction due to initial analysis. Typically, a Strategic Planning
Work Program would have the components in Table 1.
If an aspect of the project does not have a cost or revenue it will
For new greenfield or acquisition projects, approximately half
not directly be added to the project NPV, hence the inevitable
the available time is typically used to collect appropriate data,
constraints that must be considered. There are many issues to be validate it and get it into the appropriate format for use by the
considered in strategic planning that are not captured directly in planning software. Once a model is validated, a new expansion
the objective function. Safety and environmental care are two study, or the incorporation of additional reserves, requires much
priorities of modern responsible mining that have important issues less data validation and the emphasis shifts to the analysis of
that are not always evident in cost or revenue. These and many results. Several of the tasks in Table 1 are done in parallel due to
more issues are captured in the optimisation process as constraints the iterative nature of the planning process. Note that Table 1
to bound the schedules that are considered valid, and are known as does not include the training time to ensure people understand
the feasible solutions. Many processes have a combination of what to do in each of the phases.
OPTIMAL MINING PRINCIPLES
For annual plans that cover the entire life of a major mineral
TABLE 1 resource, we will not typically model individual salaries, weekly
Summary of typical strategic planning effort. maintenance schedules, shovel loading configurations or the
Tasks New model Existing model extra day that occurs in leap years. We must draw the distinction
between different planning time frames (real time, shift, weekly,
Scoping and understanding the 10% 10% monthly, quarterly and annual) and the appropriate level of detail
existing project
for the analysis.
Gathering, validating and 50% 20% As a general procedure, it is advisable to start off with a simple
transforming data
model of the entire project, then add complexity if required. Once
Model building 10% 10% you have a simple model it can be used and its results analysed.
Running scenarios and 20% 50% When a component of the cash flow calculation is found to be
analysing the results poorly approximating the real value, it can be modelled in
Reporting on results 10% 10% greater detail. For example, an initial model of a mine may start
with the simple mining cost per mass of material moved. This
model then may be broken down into components and their
Given how long data validation and analysis takes, it is critical drivers, such as drilling ($/t), blasting ($/t), loading ($/m3) and
to plan a realistic scope. ‘Dry run’ strategic planning exercises hauling ($/truck_operating_hour). The haulage component may
can help identify training requirements, holes in data require further detail by modelling the source elevation,
requirements, streamline data acquisition and align scope destination of the material and the height of the dumps and so on.
expectations. There are typically too many options to evaluate The objective is to arrive at a good estimate of the NPV for
and not enough time, so the planned activity should focus on several different alternatives. If a simple model provides a good
identifying the highest value controllable drivers. For example, estimate, then there is no need to pursue a complex model.
although blasthole initiation systems may be expensive and have Another advantage of this approach is that some analysis can be
room for reduction, they may not be as important to consider as done quickly into the project – this may enable the project scope
to be tailored to the value drivers as they are better quantified. It
shovel productivity as a driver of project value. Some value
is often surprising how simple models can estimate value so
drivers may not be controllable, such as taxation, statutory
accurately and drive the decision process appropriately.
compliance fees, award wages, cost of power and possibly even
the commodity price. Although these uncontrollable drivers are
fixed in the analysis, they should be considered in the NPV Knowledge and understanding – keys for
calculations, since they may influence decisions. discovering value
Simplifying assumptions This section of the paper aims to point people in directions to find
more value for their projects. It appears that no one person or
Large mineral resources may take many years or even decades to company has exclusive access to value creating ideas. It is my
fully exploit and have more decisions to make than there are conviction that many high value ideas will flow through people
electrons in the universe (approximately 1079). The currently who intimately understand a project, not just through the industry
available algorithms for mine design, scheduling and processing expert who comes in for a three day review. A sound knowledge of
require simplifying assumptions to find a solution and substantial the project, value drivers and optimisation assumptions are all
computational power to do any sort of partial optimisation. For important to unlocking greater value from a mining project. The
example, the Lerchs-Grossman (LG) algorithm, used in finding variety of geological resources, mining methods and philosophies
ultimate pit limits (Lerchs and Grossmann, 1965) – fundamentally, employed at major projects around the globe mean that value
this algorithm uses the assumption that we know the value of discovered in one project may not be found in another. The
every block in a model. This is somewhat difficult when we do not following suggestions are general areas that are regularly
know when the block is to be mined, what the price will be at that neglected and so may be sources of additional value.
time, what other blocks are being mined with this one, what
processes are available or whether the ideal process will have Project optimisation
capacity for the block. This is a simplification in order to estimate
the shape of the mine when completely mined. Of all the optimisation analysis that is undertaken in geology,
If the value is dependent on time or capacity-related issues, then mining, processing and marketing, it seems that very little analysis
this casts doubts as to just how ‘optimal’ the LG results are. This is spans all of these disciplines. Although the same detail cannot be
not to say that the algorithms cannot be used to guide the best applied for global project analysis as can be applied to detailed
location of the ultimate pit limits, it must however be taken into component analysis, we need to start with the big picture to make
account when interpreting the results. Phrases with ‘optimise’ and most effective use of our finite resources (such as time, people and
‘plans’ should be regarded with a healthy scepticism – we should budgets). The project analysis allows for determination of where
also realise that we can rarely claim to have the optimum plan, the most important value drivers are, allowing prioritising of
analysis effort. The project optimisation can also provide context
rather something closer to the highest value plan we could deliver
and scope for a more detailed analysis.
in the time allowed. The algorithms underlying all currently
available software tools have limitations due to their assumptions
– skilled engineers are able to use them appropriately in spite of Working out of your discipline
these limitations. Some other simplifications typically made in Concentrated and isolated analysis within geology, mining,
strategic planning include: processing, environmental, marketing and financial can easily
• mineral resource estimates accurately reflecting what mining develop silos of knowledge and understanding. While this may
and processing will deliver; start at university and a natural affinity with one’s discipline, it is
• annual equipment productivity, utilisation and availability essential for project optimisation that some people come to
estimates; understand the whole project. Many fields are interrelated – for
example, problems in the processing plant could well stem from
• predictable costs and prices over the life of the project; and fundamentally different geological ore genesis. The cut-off grade
• pretax and post-tax optimisation leading to the same of material mined has substantial implications on processing and
decisions. marketing parameters.
B KING
Multidisciplinary teams King, B, 1999. Cash flow grades – Scheduling rocks with different
throughput characteristics, in Proceedings Whittle Strategic Mine
Major mining projects are well beyond the technical skills of any Planning Conference, pp 103-109 (Whittle Programming: Melbourne).
one individual. Creating teams to look for value has often led to King, B, 2008. Optimal mining – Building the practice of optimisation in
unrealised value. The quantity of information to be absorbed and the mining industry [online]. Available from: <http://www.optimal
translated often works well with a multidisciplinary team. Teams mining.org/optimalmining/>.
would normally include geology, mining, processing, environ- Lane, K F, 1964. Choosing the optimum cut-off grade, Q Col Sch Mines,
mental and financial members and it is essential that some of the 59(4):811-829.
team members should also have a strong optimisation capability. Lane, K F, 1988. The Economic Definition of Ore – Cut-off Grades in
Well-run teams can encourage innovation and building on each Theory and Practice (Mining Journal Books: London).
other’s ideas – they are also often able to quickly discard options Lerchs, H and Grossman, I F, 1965. Optimum design of open-pit mines,
unlikely to add value. Transactions of the Canadian Institute of Mining, LXVIII:17-24.
Wooller, R, 2007. Optimising multiple operating policies for exploiting
Removing constraints complex resources – An overview of the COMET scheduler,
in Orebody Modelling and Strategic Mine Planning, second edition
Because constraints place substantial limits on the value of a (ed: R Dimitrakopoulos), pp 309-316 (The Australasian Institute of
project, what is the value of removing these constraints? The Mining and Metallurgy: Melbourne).
concept of de-bottlenecking or using constraints theory may be
used, as long as it is in conjunction with value analysis and not
just for gaining additional capacity. If the primary crusher is a APPENDIX – APPLYING BREAK-EVEN
major bottleneck, would the value obtained by adding capacity CUT-OFF GRADES
cover the new capacity? The following example has been presented to show how care must
be exercised when determining what costs should be included in
Constraints or targets calculating operational cut-off grades. This example highlights a
number of scenarios involving reclaim from stockpiles and which
The direction of a project can easily take on the preconceived costs should be included in each case. Consider the costs that are
ideas of an influential manager. Creating smooth ore-grade used to calculate a mill cut-off grade – should the mining (drill,
profiles has a cost to the mine – what is the value implication if blast, load and haul) and G&A (General & Administration, $/year)
the processing plant accepted more variability? Marketing type costs be included for material going to or coming from a
‘constraints’ may be little more than market predictions; are stockpile?
these creating ceilings on increasing the value of a project? Many
of these issues are quickly identified by multidisciplinary teams. First, take the example of a project that is ‘mine constrained’ –
the project is struggling to use the processing capacity due to
mining capacity restrictions (truck or shovel capacity, sinking
Full value focus rates, etc). This will typically occur at the beginning of a project
Focus on the corporate objective of what adds value, even if that when the high value material is only located at depth. It also may
does not suit some parties. Utilise the whole of the resource occur after mining one high grade area before the next high grade
rather than just the first 20 to 30 years (it will help demonstrate area is available. Even when using all the mining equipment, there
the value of expansions). Making an open pit bigger should be may not be sufficient high grade ore to fill the mill capacity.
considered in the light of the alternate underground options. How What is the minimum grade that can be sent for processing?
can the project realise the greatest value from one of these The question can easily be answered by considering the two
options? Don’t just use average quantities – know the value of: options possible for a region of ground:
• a truck of ore, 1. What is the value of sending the material to the mill?
• a hole of explosives, 2. What is the value of sending the material to waste (Waste
• a conveyor belt, value will be a negative number since all components are
• SAG steel balls, and negative value costs)?
• acid and wheel motors. Since the material is being mined, when the value of sending
the material to the mill is greater than sending it to the waste
The more people understand the value, the more they are likely dumps, it should be processed. Using variable definitions as used
to make appropriate value based decisions. Break-even cut-off by Lane (1988) where possible, such as the period cost (f, $/year),
grade costs calculations for reclaiming stockpiles are presented processing cost (h, $/t ore), mining cost (m, $/t rock), selling cost
in an appendix of this paper to illustrate the need and value of (k, $/t metal), price (p, $/t metal), recovery (y, recovered metal/in
understanding optimisation assumptions. situ metal), mining capacity (M, t rock/year), mass above cut-off
grade (x, t ore/t rock) and grade (g, in situ metal/t rock), the
CONCLUSIONS following relationships are clear:
This paper has provided guidance to help strategic planning VORE = (p-k).y.g.x – f/M – h.x - mORE
engineers to fully and profitably exploit the resources in their
care. Shareholder value within a constraint framework is
normally the objective for optimal mine plans – the key is to not VWASTE = -f/M - mWASTE
get so distracted by details that focus is lost on what really adds The mining cost of processing material as ore (mORE) could be
value. These are finite resources that we normally only get one
different to that when the same material is processed as waste
chance to exploit – let’s make the most of them while achieving
the best value for shareholders. (mWASTE). For example, some operations will blast the material
more coarsely if it is being sent to waste, and the haulage costs
are often different between ore and waste. The mine constrained
REFERENCES cut-off grade for material mined from the pit (gMP) is determined
Bellman, R, 1957. Dynamic Programming (Princeton University Press: when VORE = VWASTE as follows:
New Jersey).
h + mORE − mWASTE
Brealey, R A and Myers, S C, 2000. Principles of Corporate Finance, gM =
sixth edition (Irwin McGraw-Hill). ( p − k ). y
OPTIMAL MINING PRINCIPLES
Apparent from the derivations and the formula for gM above, VWASTE =0
the period costs do not have to be covered in this situation. This
observation makes sense when one considers that the period The stockpile cut-off grade (gs) at the end of the project life is
costs will be incurred irrespective of the destination of the again determined when VORE = VWASTE as follows:
material. When the ore and waste mining costs are identical, they f
also drop out of the calculation and it becomes the same equation h + mORE +
that Ken Lane formulated (1988, page 28). This shows that gS = H
material mined in a pit will incur mining costs regardless of its ( p − k ). y
final destination. The decision can simply be made depending on
This is an interesting result because now the break-even cut-off
whether the project makes more money from processing the rock
grade for processing the last material from the mine stockpile
than sending it to the waste dump. Where a stockpile exists
must include the mining costs and also the period costs. If
during the same mine constrained period, the following equation
material is added to a stockpile using mine constrained cut-off
to calculate the mine constrained stockpile break-even cut-off
grades, then there could be substantial uneconomic material
grade (gMS) simplifies to:
mixed in with the economic material and it is likely impossible
h + mORE to separate. The situation has more issues to be considered when
gMS =
( p − k ). y the material mined is not at the end of the project life and an
opportunity cost exists. There is an opportunity cost to reflect
The mORE cost for a stockpile normally only includes the which material is required to fill the limited processing capacity,
reclaim loading and haulage costs, since drilling and blasting costs and the timing to access the remaining material. For a processing
are not necessary. Notice that here the waste mining cost does not constrained stockpile, a formulation of the opportunity cost
incur any costs since the material can simply be left on the defines an economic cut-off grade (gE) as:
stockpile. The stockpile re-handling costs must be covered by the f +F
processing revenue in this case. This means that higher grade h + mORE +
material would need to be reclaimed from the stockpile than what gE = H
could be sent directly from fresh material mined in the pit. Ore ( p − k ). y
ranking grades need to be updated (such as the cash flow grades,
In practice, this formulation is very simplistic since it needs to
King, 1999) to take this into account as material is moved from the
take into account multiple sequences, several processes,
pit into a stockpile.
interaction between several ore processing constraints, revenue
A subtle assumption in the above cut-off grade derivations is from multiple elements and time varying issues. It does show
that the material will be processed. Let’s consider a stockpile at that there is not one simply universal answer for the inclusion of
the end of the project life. If the material is not processed, it is period and mining costs in determining what material should be
left on the ground and the project abandoned (ignoring closure mined from, or deposited to, a stockpile. By taking the time to
costs and reclamation for the moment). It is also assumed that understand the project situation and algorithm assumptions,
the environment is no longer mine constrained, but rather we are appropriate material can be sent to stockpiles and reclaimed.
processing constrained (H, t ore/year), and all the material mined Although opportunity costs or remaining value estimates are
from the stockpile is ore (x = 1). In this case we consider the implemented in modern cut-off grade optimisation software tools
following equations: like Comet (Wooller, 2007), they may not find the additional
VORE = (p-k).y.g – f/M – h - mORE value without a model constructed to look for this value.
HOME
Stochastic Mine Planning – Methods, Examples and Value in an

Uncertain World
R Dimitrakopoulos1
ABSTRACT
Conventional approaches to estimating reserves, optimising mine
planning and production forecasting result in single, often biased,
forecasts. This is largely due to the non-linear propagation of errors in
understanding orebodies throughout the chain of mining. A new mine
planning paradigm is considered herein, integrating two elements:
stochastic simulation and stochastic optimisation. These elements provide
an extended mathematical framework that allows modelling and direct
integration of orebody uncertainty to mine design, production planning,
and valuation of mining projects and operations. This stochastic
framework increases the value of production schedules by 25 per cent.
Case studies also show that stochastic optimal pit limits:
• can be about 15 per cent larger in terms of total tonnage when
compared to the conventional optimal pit limits, while
• adding about 10 per cent of net present value (NPV) to that reported
above for stochastic production scheduling within the conventionally
optimal pit limits.
Results suggest a potential new contribution to the sustainable FIG 1 - Optimisation of mine design in an open pit gold mine, net
utilisation of natural resources. present value versus ‘pit shells’ and risk profile of the
conventionally optimal design.
INTRODUCTION
Optimisation is a key aspect of mine design and production model as an input to optimisation for mine design and a ‘correct’
scheduling for both open pit and underground mines. It deals with assessment of individual key project indicators, a set of models
the forecasting, maximisation, and management of cash flows of the deposit can be used. These models are conditional to the
from a mining operation and is the key to the financial aspects of same available data and their statistical characteristics, and all are
mining ventures. A starting point for optimisation in the above constrained to reproduce all available information and represent
context is the representation of a mineral deposit in equally probable models of the actual spatial distribution of grades
three-dimensional space through an orebody model and the (Journel, 1994). The availability of multiple equally probable
mining blocks representing it; this is used to optimise designs and models of a deposit enables mine planners to assess the sensitivity
production schedules (eg Whittle, 1999). Geostatistical estimation of pit design and long-term production scheduling to geological
methods have long been used to model the spatial distribution of uncertainty (eg Kent, Peattie and Chamberlain, 2007; Godoy,
grades and other attributes of interest within the mining blocks 2010, in this volume) and, more importantly, empower mine
representing a deposit (David, 1988). The main drawback of planners to produce mine designs and production schedules with
estimation techniques, be they geostatistical or not, is that they are substantially higher NPV assessments through stochastic
unable to reproduce the in situ variability of the deposit grades, as optimisation. Figure 3 shows an example from a major gold mine
inferred from the available data. Ignoring such a consequential presented in Godoy and Dimitrakopoulos (2004), where a
source of risk and uncertainty may lead to unrealistic production stochastic approach leads to a marked improvement of 28 per
expectations (eg Dimitrakopoulos, Farrelly and Godoy, 2002). cent in NPV over the life of the mine, compared to the standard
Figure 1 shows an example of unrealistic expectations in a best practices employed at the mine; note that the pit limits used
relatively small gold deposit. In this example (Dimitrakopoulos, are the same in both cases and are conventionally derived
Farrelly and Godoy, 2002), the smoothing effect of estimation through commercial optimisers (Whittle, 1999). The same study
methods generates unrealistic expectations of net present value in also shows that the stochastic approach leads to substantially
the mine’s design, along with ore production performance, pit lower potential deviation from production targets, that is, reduced
limits, and so on. The figure shows that if the conventionally risk. A key contributor to substantial differences is that the
constructed open pit design is tested against equally probable stochastic or risk-integrating approach can distinguish between
simulated scenarios of the orebody, its performance will probably the ‘upside potential’ of the metal content, and thus economic
not meet expectations. The conventionally expected NPV of the value of a mining block, from its ‘downside risk’, and then treat
mine has a 2 - 4 per cent chance to materialise, while it is expected them accordingly, as further discussed herein.
to be about 25 per cent less than forecasted. Note that in a Figure 2 represents an extended mine planning framework that
different example, the opposite could be the case. is stochastic (that is, integrates uncertainty) and encompasses the
For over a decade now, a traditional framework has been used spatial stochastic model of geostatistics with that of stochastic
when dealing with uncertainty in the spatial distribution of optimisation for mine design and production scheduling. Simply
attributes of a mineral deposit, as well as its implications to put, in a stochastic mathematical programming model developed
downstream studies, planning, valuation, and decision-making. for mine optimisation, the related coefficients are correlated
Now, a different framework than the traditional has been random variables that represent the economic value of each block
suggested and is outlined in Figure 2. Instead of a single orebody being mined in a deposit, which are in turn generated from
considering different realisations of metal content. Note that the
1. FAusIMM, Professor and Director, COSMO – Stochastic Mine
second key element of the risk-integrating approaches is
Planning Laboratory, Department of Mining and Materials stochastic simulation; the reader is referred to Mustapha and
Engineering, McGill University, Montreal QC H3A 2A7, Canada. Dimitrakopoulos (2010, in this volume) for the description of
Email: [email protected] a new general method for high-order simulation of complex
R DIMITRAKOPOULOS
FIG 2 - Traditional (deterministic or single model) view and practice versus risk-integrating (or stochastic) approach to mine modelling, from
reserves to production planning and life-of-mine scheduling, and assessment of key project indicators.
STOCHASTIC OPTIMISATION IN MINE DESIGN

AND PRODUCTION SCHEDULING
Mine design and production scheduling for open pit mines is an
intricate, complex, and difficult problem to address due to its
large-scale and uncertainty in the key parameters involved. The
objective of the related optimisation process is to maximise the
total net present value of the mine plan. One of the most
significant parameters affecting the optimisation is the
uncertainty in the mineralised materials (resources) available in
the ground, which constitutes an uncertain supply for mine
production scheduling. A set of simulated orebodies provides a
quantified description of the uncertain supply. Two stochastic
optimisation methods are summarised in this section. The first is
FIG 3 - The stochastic life-of-mine schedule in this large gold mine
based on simulated annealing (Godoy and Dimitrakopoulos,
has a 28 per cent higher value than the best conventional
(deterministic) one. All schedules are feasible.
2004; Leite and Dimitrakopoulos, 2007; Albor Consequega and
Dimitrakopoulos, 2009); and the second on stochastic integer
programming (Ramazan and Dimitrakopoulos, 2007, 2008;
geological phenomena. The further integration of market Menabde et al, 2007; Leite and Dimitrakopoulos, 2010, in this
uncertainties in terms of commodity prices and exchange rates is volume).
discussed elsewhere (Abdel Sabour and Dimitrakopoulos, 2010,
in this volume; Meagher, Abdel Sabour and Dimitrakopoulos,
2010, in this volume). Production scheduling with simulated annealing
The key idea in production scheduling that accounts for grade Simulated annealing is a heuristic optimisation method that
uncertainty is relatively simple. A conventional optimiser (any integrates the iterative improvement philosophy of the so-called
one of them) is deterministic by construction and evaluates a Metropolis algorithm with an adaptive ‘divide and conquer’
cluster of blocks, such as that in Figure 4a, so as to decide when strategy for problem solving (Geman and Geman, 1984). When
to stop mining, which blocks to extract when, and so on,
several mine production schedules are under study, there is
assuming that the economic values of the mining blocks
considered (inputs to the optimiser) are the actual/real values. A always a set of blocks that are assigned to the same production
stochastic optimiser, also by construction, evaluates a cluster of period throughout all production schedules; these are referred to
blocks, but as in Figure 4b, by simultaneously using all possible as the certain or 100 per cent probability blocks. To handle the
combinations of economic values of the mining blocks in the uncertainty in the blocks that do not have 100 per cent
cluster being considered. As a result, substantially more local probability, simulated annealing swaps these blocks between
information on joint local uncertainty is utilised, leading to much candidate production periods so as to minimise the average
more robust schedules that also can maximise the upside deviation from the production targets for N mining periods, and
potential of the deposit (eg higher NPV and metal production) for a series of S simulated orebody models, that is:
and at the same time minimise downsides (eg not meeting
production targets and related losses). N
⎛ S S
⎞
To elaborate on the above, the next sections examine a key MinO = ∑ ⎜ ∑ θ *n (s) − θ n (s) + ∑ ω *n (s) − ω n (s)⎟ , (1)
n =1 ⎝ s=1 s=1 ⎠
element in the risk-integrating framework shown in Figure 2, that
of stochastic optimisation. The latter optimisation is presented in
two approaches, one based on the technique of simulated where θ *n ( s ) and ω *n ( s ) are the ore and waste production targets,
annealing, and a second based on stochastic integer programming. respectively, θ n ( s ) and ω n ( s ) represent the actual ore and waste
Examples follow that demonstrate the practical aspects of production of the perturbed mining sequence. Each swap of a
stochastic mine modelling, including the monetary benefits. block is referred to as a perturbation.
STOCHASTIC MINE PLANNING – METHODS, EXAMPLES AND VALUE IN AN UNCERTAIN WORLD
FIG 4 - Production scheduling optimisation with conventional versus stochastic optimisers: (a) single representation of a cluster of mining
blocks in a mineral deposit as considered for scheduling by a conventional optimiser; and (b) a set of models of the same cluster of blocks
with multiple possible values considered simultaneously for scheduling by a stochastic optimiser.
The probability of acceptance or rejection of a perturbation is models as input, without averaging the related grades. The optimal
given by: production schedule is then the schedule that can produce the
maximum achievable discounted total value from the project,
⎧1, if Onew ≤ Oold given the available orebody uncertainty described through a set of
⎪
Prob{accept} = ⎨ Oold − Onew . stochastically simulated orebody models. The proposed SIP model
⎪⎩e T
, otherwise allows the management of geological risk in terms of not meeting
planned targets during actual operation. This is unlike the
This implies all favourable perturbations (Onew ≤ Οold) are traditional scheduling methods that use a single orebody model,
accepted with probability 1 and unfavourable perturbations are and where risk is randomly distributed between production periods
accepted based on an exponential probability distribution, where while there is no control over the magnitude of the risks on the
T represents the annealing temperature. schedule.
The steps of this approach, as depicted in Figure 5 are as The general form of the objective function is expressed as:
follows:
⎡ n ⎤
( )
p m
MAX ∑ ⎢ ∑ E ⎧⎨( NPV) ⎫⎬ b ti − ∑ c tou d tosu + c to1 d tos1 + c tgu d tgsu + c tg1 d tgs1 ⎥,
t
1. define ore and waste mining rates;
t =1 ⎣ i = 1 ⎩ i
⎭ s=1 ⎦
2. define a set of nested pits as per the Whittle implementation
(Whittle, 1999) of the Lerchs-Grossmann (1965) algorithm, (2)
or any pit parameterisation;
where:
3. use a commercial scheduler to schedule a number of
simulated realisations of the orebody given 1 and 2; p is the total production periods
4. employ simulated annealing as in Equation 1 using the n is the number of blocks
results from 3 and a set of simulated orebodies; and bi t is the decision variable for when to mine block i (if mined
5. quantify the risk in the resulting schedule and key project in period t, bit is 1 and otherwise bit is 0)
indicators using simulations of the related orebody. The c variables are the unit costs of deviation (represented by
the d variables) from production targets for grades and ore
Stochastic integer programming for mine tonnes. The subscripts u and l correspond to the deviations and
production scheduling costs from excess production (upper bound) and shortage in
Stochastic integer programming (SIP) provides a framework for production (lower bound), respectively, while s is the simulated
optimising mine production scheduling considering uncertainty orebody model number, and g and o are grade and ore production
(Dimitrakopoulos and Ramazan, 2008). A specific SIP targets. Figure 6 graphically shows the second term in Equation 2.
formulation is briefly shown here that generates the optimal Note that the cost parameters in Equation 2 are discounted by
production schedule using equally probable simulated orebody time using the geological risk discount factor developed in
R DIMITRAKOPOULOS
FIG 5 - Steps needed for the stochastic production scheduling with simulated annealing. S1... Sn are realisations of the orebody grade
through a sequential simulation algorithm. Seq1…Seqn are the mining sequences for each of S1…Sn. Mining rates are input to the process.
EXAMPLES AND VALUE OF THE STOCHASTIC

FRAMEWORK
The example discussed herein shows long-range production
scheduling with both the simulating annealing approach in
Section 3.1 and SIP model in Section 3.2. Section 3.3 focuses on
the topic of stochastically optimal pit limits. The application
used is at a copper deposit comprising 14 480 mining blocks.
The scheduling considers an ore capacity of 7.5 M tonnes per
year and a maximum mining capacity of 28 M tonnes. All results
are compared to the industry’s ‘best practice’: a conventional
schedule using a single estimated orebody model and Whittle’s
approach (Whittle, 1999).
Simulated annealing and production schedules

The results for simulated annealing and the method in Equation 1
are summarised in Figures 7 to 10. The risk profiles for NPV, ore
tonnages, and waste production are respectively shown in
FIG 6 - Graphic representation of the way the second component of Figures 7, 8, and 9. Figure 10 compares with the equivalent best
the objective function in Equation 2 minimises the deviations from conventional practice and reports a difference of 25 per cent in
production targets while optimising scheduling. This leads to terms of higher NPV for the stochastic approach.
schedules where the potential deviations from production targets
are minimised, leading to schedules that seek to mine first not only
for high-grade mining blocks, but also with high probability to be ore.
Dimitrakopoulos and Ramazan (2004). The geological risk

discount rate (GRD) allows the management of risk to be
distributed between periods. If a very high GRD is used, the
lowest risk areas in terms of meeting production targets will be
mined earlier and the most risky parts will be left for later
periods. If a very small GRD or a GRD of zero is used, the risk
will be distributed at a more balanced rate among production
periods depending on the distribution of uncertainty within the
mineralised deposit. The ‘c’ variables in the objective function
(Equation 2) are used to define a risk profile for the production,
and NPV produced is the optimum for the defined risk profile. It
is considered that if the expected deviations from the planned
amount of ore tonnage having planned grade and quality in a
schedule are high in actual mining operations, it is unlikely to
achieve the resultant NPV of the planned schedule. Therefore,
the SIP model contains the minimisation of the deviations
together with the NPV maximisation to generate practical and
feasible schedules and achievable cash flows. For details, please
see Ramazan and Dimitrakopoulos (2008) and Dimitrakopoulos FIG 7 - Risk based life-of-mine production schedule (cumulative net
and Ramazan (2008). present value risk profile).
Stochastic integer programming and production

schedules
The application of the SIP model in Equation 2, using pit limits
derived from the conventional optimisation approach, forecasts
an expected NPV at about $238 M. When compared to the
equivalent traditional approach and related forecast, the value of
the stochastic framework is $60 M, or a contribution of about
25 per cent additional NPV to the project. Note that unlike
simulated annealing, the scheduler decides the optimal waste
removal strategy, which is the same as the one used in the
conventional optimisation with which we compare.
Figure 11 shows a cross-section of the two schedules from the
copper deposit: one obtained using the SIP model (bottom) and
the other generated by a traditional method (top) using a single
estimated orebody model. Both schedules shown are the raw
outputs and need to be smoothed to become practical. It is
important to note that:
• the results in the second case study are similar in a
percentage improvement when compared to other stochastic
FIG 8 - Risk based life-of-mine production schedule (ore risk profile). approaches such as simulated annealing; and
• although the schedules compared in the studies herein are not
smoothed out, other existing SIP applications show that the
effect of generating smooth and practical schedules has
marginal impact on the forecasted performance of the related
schedules, thus the order of improvements in SIP schedules
reported here remains.
Stochastically optimal pit limits

The previous comparisons were based on the same pit limits
deemed optimal using best industry practice (Whittle, 1999).
This section focuses on the value of the proposed approaches
with respect to stochastically optimal pit limits. Both methods
described above consider larger pit limits and stop when
discounted cash flows are no longer positive. Figures 12 and 13
show some of the results. The stochastically generated optimal
pit limits contain an additional 15 per cent of tonnage when
compared to the traditional (deterministic) ‘optimal’ pit limits,
add about 10 per cent in NPV to the NPV reported above from
stochastic production scheduling within the conventionally
optimal pit limits, and extend the life-of-mine. These are
substantial differences for a mine of a relatively small size and
short life-of-mine. Further work shows that there are additional
FIG 9 - Risk based life-of-mine production schedule (waste risk improvements on all aspects when a stochastic framework is
profile). used for mine design and production scheduling.
FIG 10 - Net present value of conventional and stochastic (risk based) schedules and corresponding risk profiles.
R DIMITRAKOPOULOS
CONCLUSIONS
Mining Periods Starting from the limits of the current orebody modelling and
1 life-of-mine planning optimisation paradigm, an integrated
2 risk-based framework has been presented. This framework
3 extends the common approaches in order to integrate both
4 stochastic modelling of orebodies and stochastic optimisation in
5 a complementary manner. The main drawback of estimation
6 techniques and traditional approaches to planning is that they are
7
8 unable to account for the in situ spatial variability of the deposit
grades; in fact, conventional optimisers assume perfect knowledge
of the orebody being considered. Ignoring this key source of risk
and uncertainty can lead to unrealistic production expectations as
well as suboptimal mine designs.
The work presented herein shows that the stochastic
framework adds higher value in production schedules in the
order of 25 per cent, and will be achieved regardless of which
method from the two presented is used. Furthermore, stochastic
optimal pit limits are shown to be about 15 per cent larger in
terms of total tonnage, compared to the traditional (deterministic)
FIG 11 - Cross-sectional views of the Stochastic integer optimal pit limits. This difference extends the life-of-mine and
programming (bottom) and traditional schedule (top) for a copper adds approximately ten per cent of net present value (NPV) to the
deposit. NPV reported above from stochastic production scheduling within
the conventionally optimal pit limits.
The new approach yielded an increment of ~30 per cent in the
NPV when compared to the conventional approach. The ACKNOWLEDGEMENTS
differences reported are due to the different scheduling patterns,
the waste mining rate, and an extension of the pit limits which Thanks are in order to the International Association of
yielded an additional ~5.5 thousand tonnes of metal. Mathematical Geosciences for the opportunity to present this
FIG 12 - Life-of-mine cumulative cash flows for the conventional approach, simulated annealing and SIP, compared to results
from conventionally derived optimal pit limits.
FIG 13 - Stochastic pit limits are larger than the conventional ones; physical scheduling differences are expected when bigger pits
are generated.
work as their distinguished lecturer. The support of the COSMO Kent, M, Peattie, R and Chamberlain, V, 2007. Incorporating grade
Laboratory and its industry members AngloGold Ashanti, uncertainty in the decision to expand the main pit at the Navachab
Barrick, BHP Billiton, De Beers, Newmont, Vale and Vale Inco, gold mine, Namibia, through the use of stochastic simulation, in
Orebody Modelling and Strategic Mine Planning, second edition (ed:
as well as NSERC, the Canada Research Chairs Program and R Dimitrakopoulos), pp 207-216 (The Australasian Institute of
CFI is gratefully acknowledged. Thanks to R Goodfellow for Mining and Metallurgy: Melbourne).
editorial assistance. Leite, A and Dimitrakopoulos, R, 2007. A stochastic optimization model
for open pit mine planning: Application and risk analysis at a copper
REFERENCES deposit, Transactions of the Institutions of Mining and Metallurgy,
Mining Technology, 116(3):A109-A118.
Abdel Sabour, S A and Dimitrakopoulos, R, 2010. Accounting for joint Leite, A and Dimitrakopoulos, R, 2010. Production scheduling under
ore supply, metal price and exchange rate uncertainties in mine metal uncertainty – Application of stochastic mathematical
design, in Advances in Orebody Modelling and Strategic Mine programming at an open pit copper mine and comparison to
Planning I (ed: R Dimitrakopoulos), pp 29-34 (The Australasian conventional scheduling, in Advances in Orebody Modelling and
Institute of Mining and Metallurgy: Melbourne). Strategic Mine Planning I (ed: R Dimitrakopoulos), pp 35-40 (The
Albor Consquega, F and Dimitrakopoulos, R, 2009. Stochastic mine Australasian Institute of Mining and Metallurgy: Melbourne).
design optimization based on simulated annealing: Pit limits, Lerchs, H and Grossmann, I F, 1965. Optimum design of open-pit mines,
production schedules, multiple orebody scenarios and sensitivity Transactions of the Canadian Institute of Mining and Metallurgy,
analysis, Transactions of the Institutions of Mining and Metallurgy, LXVII:47-54.
Meagher, C, Abdel Sabour, S A and Dimitrakopoulos, R, 2010. Pushback
David, M, 1988. Handbook of Applied Advanced Geostatistical Ore design of open pit mines under geological and market uncertainties,
Reserve Estimation, 216 p (Elsevier Science: Amsterdam). in Advances in Orebody Modelling and Strategic Mine Planning I
Dimitrakopoulos, R, Farrelly C T and Godoy, M, 2002. Moving forward (ed: R Dimitrakopoulos), pp 291-298 (The Australasian Institute of
from traditional optimization: Grade uncertainty and risk effects in Mining and Metallurgy: Melbourne).
open-pit design, Transactions of the Institutions of Mining and Menabde, M, Froyland, G, Stone, P and Yeates, G, 2007. Mining
Metallurgy, Mining Technology, 111:A82-A88. schedule optimisation for conditionally simulated orebodies, in
Dimitrakopoulos, R and Ramazan, S, 2004. Uncertainty based production Orebody Modelling and Strategic Mine Planning, second edition (ed:
scheduling in open pit mining, SME Transactions, 316:106-112. R Dimitrakopoulos), pp 379-384 (The Australasian Institute of
Dimitrakopoulos, R, and Ramazan, S, 2008. Stochastic integer Mining and Metallurgy: Melbourne).
programming for optimizing long term production schedules of open Mustapha, H and Dimitrakopoulos, R, 2010. Geologically enhanced
pit mines: Methods, application and value of stochastic solutions, simulation of complex mineral deposits through high-order spatial
Transactions of the Institutions of Mining and Metallurgy, Mining cumulants, in Advances in Orebody Modelling and Strategic Mine
Technology, 117(4):A155-A167. Planning I (ed: R Dimitrakopoulos), pp 309-320 (The Australasian
Geman, S and Geman, D, 1984. Stochastic relaxation, Gibbs distribution, Institute of Mining and Metallurgy: Melbourne).
and the Bayesian restoration of images, IEEE Transactions on Ramazan, S and Dimitrakopoulos, R, 2007. Stochastic optimisation of
Pattern Analysis and Machine Intelligence, PAMI-6(6):721-741. long-term production scheduling for open pit mines with a new
Godoy, M, 2010. A risk analysis based framework for strategic mine integer programming formulation, in Orebody Modelling and
planning and design – Method and application, in Advances in Strategic Mine Planning, second edition (ed: R Dimitrakopoulos),
Orebody Modelling and Strategic Mine Planning I (ed: R pp 385-392 (The Australasian Institute of Mining and Metallurgy:
Dimitrakopoulos), pp 21-28 (The Australasian Institute of Mining Melbourne).
and Metallurgy: Melbourne). Ramazan, S and Dimitrakopoulos, R, 2008. Production scheduling with
Godoy, M C and Dimitrakopoulos, R, 2004. Managing risk and waste uncertain supply – A new solution to the open pit mining, COSMO
mining in long-term production scheduling, SME Transactions, research report, number 2, pp 257-294.
316:43-50. Whittle, J, 1999. A decade of open pit mine planning and optimisation –
Journel, A G, 1994. Modelling uncertainty: Some conceptual thoughts, in The craft of turning algorithms into packages, in Proceedings
Geostatistics for the Next Century (ed: R Dimitrakopoulos) (Kluwer APCOM ’99, Computer Applications in the Minerals Industries:
Academic: Dordrecht). 28 International Symposium, pp 15-24 (Colorado School of Mines:
Golden).
HOME
A Risk Analysis Based Framework for Strategic Mine Planning

and Design – Method and Application
M Godoy1
ABSTRACT UNCERTAINTY ANALYSIS FOR AN ULTIMATE

Assessment and management of orebody uncertainty is critical to PIT SHELL
strategic mine planning. This paper presents an approach that consists of The objective of this analysis is to evaluate the sensitivity of key
a series of procedures for risk assessment in pit optimisation and design. pit optimisation results to grade uncertainty. The process known as
Multiple block grade simulations are processed in Whittle Software to
pit optimisation is traditionally carried out based on the estimated
produce a distribution of possible outcomes in terms of net present value.
Examples from an open pit mine are used to illustrate the practical
resource model and using the nested pit implementation of the
application of the methodology. Lerchs-Grossmann algorithm of the Whittle Software (Lerchs and
Grossmann, 1965; Whittle, 1999). The result of the optimisation
process is a series of incremental pit shells. Different criteria can
INTRODUCTION be used to select the ultimate pit shell including net pit value and
the net present value based on a referential mining sequence. This
Traditionally, determination of the spatial distribution of grades ultimate pit shell is then used as the basis for pit design and
in an orebody model is based on geostatistical estimation. The planning. Using conditionally simulated models as input, Whittle’s
main drawback of estimation techniques, be they geostatistical or analysis program (FDAN) may be used allowing the quantitative
not, is that they are unable to reproduce the in situ spatial assessment of risk due to uncertainty on the real, but unknown,
distribution of grades. In the procedure proposed below an
variability as inferred from the available data. Ignoring such a
ultimate pit shell produced in Whittle is evaluated against a series
consequential source of risk and uncertainty may lead to of simulated models of the orebody.
unrealistic production plans (eg Dimitrakopoulos, Farrelly and
Godoy, 2002). In dealing with uncertainty on the spatial To assess uncertainty on the main parameters driving the
selection of the ultimate pit shell the following procedure is
distribution of attributes of a mineral deposit, several models of
proposed:
the deposit can be generated based on and conditional to the
same available data and their statistical characteristics. These • From the Whittle result file produced by the pit optimisation
models are all constrained to: process, generate a Whittle pit list file containing information
about the smallest numbered pit that each block is part of.
• reproduce all available information and their statistics, and
• Apply the pit list file produced in Step 1 to each one of the
• represent equally probable models of the actual spatial simulated orebody models.
distribution of grades.
• Run the analysis program configured to generate the same
The availability of multiple equally probable models of a mine information, previously generated by analysis on the original
deposit enables mine planners to assess the sensitivity of pit pit optimisation. The analysis must be carried out for each
design and long-term production scheduling to geological one of the available simulated models.
uncertainty. This approach has been proposed by many authors
over the last 20 years (David, 1988; Journel, 1992; Ravenscroft, The above procedure generates a range of alternative outcomes
for the original optimisation process. This allows the planner to
1992; Dimitrakopoulos, 1998; Kent, Peattie and Chamberlain,
evaluate the likely range of contained ore, metal and a series
2007; and others). Figure 1 illustrates the difference between the
of key economic indicators. Figure 2 shows the predicted ore
traditional process used to convert a mineral resource into an ore tonnage produced by risk analysis. Up until pit 27, the cloud of
reserve and the risk based approach based on the technique of cumulative tonnages derived from the simulated models present
conditional simulation. an average decrease of approximately 9.76 per cent in relation to
The goal of this paper is to provide mining planning engineers the tonnage predicted by the estimated model. At pit 27 the
a series of procedures that can be used to consider the effects of estimated model indicates approximately 180.3 Mt against an
grade uncertainty in mine planning studies. Four cases have been expected value of 164.6 Mt derived from the simulations. The
selected to illustrate different types of applications: expected outcomes of contained ore go from 163.3 Mt to
1. uncertainty analysis of an ultimate pit shell – Net Value 166.1 Mt, which corresponds to a range of -0.83 per cent to
+0.92 per cent in relation to the expected value derived from the
(NV), costs, tonnage, grade and metal;
simulations. The same type of analysis carried out on the average
2. identification of areas of upside potential and downside mill feed grade (Figure 3) shows a decreasing overestimation in
risk; grades of the estimated model in relation to the simulated
models. At pit shell number 1, the estimated model predicts an
3. trade-off analysis for cut-back depletion strategies; and average mill feed grade of 2.6 g/t, while the simulations indicate
4. assessment of uncertainty related to ore blocks driving the an expected grade of 2.2 g/t. At pit 27 the estimated model
increment between the successive pit shells. indicates approximately 1.98 g/t against an expected value of
1.95 g/t derived from the simulations. The risk profile on mill
In the following sections, each application is developed
feed grade goes from 1.93 g/t to 1.98 g/t, which corresponds to a
separately and includes a step-by-step description of the
range of -1.24 per cent to +1.32 per cent in relation to the
procedure and a discussion of the results obtained in a real case
expected value. Figure 4 shows the predicted recovered metal
application.
produced by the analysis. Up until pit 27, the cloud of cumulative
recovered ounces derived from the simulated models present an
1. MAusIMM, Expert Mining Engineer/Geostatistician, Golder average decrease of approximately 17.41 per cent in relation to
Associates, Av 11 de Septiembre 2353, Piso 2, Proviendencia the ounces predicted by the estimated model. This result
Santiago, Chile. Email: [email protected] indicates that the estimated model is potentially overestimating
M GODOY
1
Single,
Unknown, often
Probability
true precise,
answer wrong
Single estimated
Traditional view answer
model
Reserves
Mining Process or
Transfer Function
Financial and
Mine Design &
Orebody Model Production
Production Scheduling
Forecasts
Multiple probable Accurate

1
models uncertainty
estimation
Probability
Risk oriented view
Reserves
FIG 1 - Conversion of a mineral resource into an ore reserve, traditional and risk oriented views.
Ore Quantity
350 Recovered Metal
18000
300
16000
250
14000
Ore (Mt)
200 12000
Gold (Moz)
ResSim1 ResSim2 10000

150 ResSim3 ResSim4
ResSim5 ResSim6 8000 ResSim1 ResSim2
ResSim7 ResSim8 ResSim3 ResSim4
100
ResSim9 ResSim10 6000 ResSim5 ResSim6
ResSim7 ResSim8
ResSim11 ResSim12
4000 ResSim9 ResSim10
ResSim11 ResSim12
ResSim15 EstMod
0 ResSim15 EstMod
1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 45 47 49 51 0
1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 45 47 49 51
Pit Number
Pit Number
FIG 2 - Uncertainty in ore tonnes for incremental pit shells.

FIG 4 - Uncertainty in recovered ounces for incremental pit shells.
Mill Feed Grade
3
ResSim1 ResSim2
2.8 ResSim3 ResSim4 Undiscounted Cash Flow (Net Value)
ResSim5 ResSim6
2.6 ResSim7 ResSim8
1400
ResSim9 ResSim10
2.4 ResSim11 ResSim12
ResSim13 ResSim14 1200
Grade (g/t)
2.2 ResSim15 EstMod
2 1000
Net Value (M $)
1.8 800
1.6
600
1.4
1.2 400 ResSim1 ResSim2 ResSim3 ResSim4

1 ResSim9 ResSim10 ResSim11 ResSim12
200
1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 45 47 49 51 ResSim13 ResSim14 ResSim15 EstMod
Pit Number 0
1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 45 47 49 51
FIG 3 - Uncertainty in mill feed grade for incremental pit shells. Pit Number
grades as an effect of excessive smoothing. At pit 27 the FIG 5 - Uncertainty in Net Value for incremental pit shells.
estimated model indicates 9.84 Moz against an expected value of
8.85 Moz derived from the simulations. The risk profile on an average decrease of approximately 33.39 per cent in relation
recovered ounces goes from 8.72 Moz to 8.98 Moz, which to the Net Value predicted by the estimated model. It also
corresponds to a range of -1.4 per cent to +1.58 per cent in indicates a downside potential for pit 27 with expected NV
relation to the expected value. equivalent to $951.6 M or a 26.12 per cent decrease in relation to
The combination of the overestimation in ore tonnage and mill the $1288 M obtained for the analysis on the estimated block
feed grade has a direct impact on the performance of the pit by model. It is interesting to note that if Net Value was to be used as
pit Net Value. Figure 5 presents the results obtained for the pit the criterion for the selection of the ultimate pit shell the
Net Value. simulations would agree with the estimated model by indicating
It suggests that the estimated model overestimates the pit value pit 27. The range of expected Net Values goes from $890 M to
over all optimised pit shells. Up until pit 27, the cloud of $1013 M, which corresponds to -6.47 per cent to +6.41 per cent
cumulative Net Value derived from the simulated models present in relation to the expected value.
A RISK ANALYSIS BASED FRAMEWORK FOR STRATEGIC MINE PLANNING AND DESIGN – METHOD AND APPLICATION
IDENTIFICATION OF UPSIDE AND DOWNSIDE estimated and simulated pits present an equivalent quantity of
contained rock, that is, they have approximately the same
POTENTIAL volume. From around pit shell 25, there is a clear separation with
The goal of this analysis is to explore the possible downside/ the simulated pits showing a progressive increase in comparison
upside potential of the selected ultimate pit shell regarding the to the estimated pit. This scenario remains the same until pit
available grade uncertainty models. To achieve this independent shell 39, when the estimated pit starts to converge to the cloud of
pit optimisation runs are carried out on each simulated model. simulated pits. Pit shell 27 is of particular interest as it
The analysis is divided into two parts: corresponds to the pit selected as a basis for pit design. For pit
shell 27, the average contained rock over the simulated pits is
1. First, each optimisation output is evaluated in terms of 925 Mt against 828 Mt in the estimated pit, which corresponds to
contained ore, grade, metal and pit value. This provides a an increase of approximately 11.7 per cent. Figure 7 presents the
quantification of the project potential given realistic results obtained for the total contained ore. In this case,
scenarios of the spatial distribution of grades. estimated and simulated pits show similar behaviour as the total
2. The second part looks into the spatial extends of a specific rock, only that the magnitude of the differences is smaller. For pit
pit shell as produced for each independent optimisation. shell 27, the average contained ore over the simulated pits
The comparison of this ‘cloud’ of pit shells against a given is 183.9 Mt against 180.3 Mt in the estimated pit, which
pit design provides an assessment of areas with corresponds to an increase of approximately two per cent. The
upside/downside potential and can be used to define targets risk profile on the contained ore goes from 174 Mt to 195 Mt,
for additional drilling. It also provides an indication on the which corresponds to a range of -5.92 per cent to +6.06 per cent
robustness of a given pit design in relation to grade in relation to the expected value. In terms of mill feed grade, the
uncertainty. estimated pit starts with 2.61 g/t against an average 2.43 g/t over
the simulated pits. This difference decreases with incremental pit
The procedure proposed for the development of this analysis is shells and become equivalent at pit shell 27 (Figure 8).
the following: The risk profile on mill feed grade goes from 1.95 g/t to
• Create a new project in Whittle and import each one of the 2.02 g/t, which corresponds to a range of -1.61 per cent to
simulated models. Next, set the optimisation parameters for +1.92 per cent in relation to the expected value. Figure 9 and
the optimisation run. Figure 10 present the results obtained for recovered metal and
total pit value.
• Apply the same parameters to each model and run the
optimisation. The process will generate a series of Whittle As expected, the recovered metal for estimated and simulated
result files, one for each simulated model. This step usually pits follow the same trends seen on the ore tonnage graphs. For
requires a large amount of disk space. pit shell 27, the recovered gold over simulated pits is 10.03 Moz
• Run analysis program configured to generate the relevant
summary information. The analysis must be carried out for Ore Quantity
350
each one of the available simulated models.
• Produce cross-sections for a selected pit number over all 300
optimised models. 250
Contrary to developing a risk analysis on a given ultimate pit

Ore (Mt)
200
shells, as carried out in the previous section, the above procedure ResSim1 ResSim2
150
generates alternative sets of incremental shells, one set for each ResSim3
ResSim5
ResSim4
ResSim6
simulated model. Figure 6 shows the total rock contained on 100 ResSim7 ResSim8
ResSim9 ResSim10
each incremental shell as produced by each independent ResSim11 ResSim12
optimisation. ResSim15 EstMod
The thick orange line corresponds to the results obtained from 0

1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 45 47 49 51
the analysis on the incremental shells optimised on the estimated
Pit Number
model, hereafter termed estimated pit. The thin lines correspond
to the results obtained from the analysis on the incremental shells
optimised on each one of the simulated models, hereafter termed FIG 7 - Total ore tonnage contained on incremental pit shells.
simulated pits. The figure shows that up to pit shell 24 the Estimated pit is depicted by the thick orange line and the
remaining coloured thin lines correspond to simulated pits.
Total Movement
2500 Gold Grade
3
ResSim3 ResSim4 2.8 ResSim3 ResSim4
ResSim7 ResSim8 2.6
ResSim11 ResSim12 2.4 ResSim11 ResSim12

Rock (Mt)
ResSim13 ResSim14
Grade (g/t)
2.2 ResSim15 EstMod

ResSim15 EstMod
2
1000
1.8
1.6
500
1.4
1.2
0
1
1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 45 47 49 51
1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 45 47 49 51
Pit Number
Pit Number
FIG 6 - Total rock tonnage contained on incremental pit shells. FIG 8 - Average grade contained on incremental pit shells.
Estimated pit is depicted by the thick orange line and the Estimated pit is depicted by the thick orange line and the
remaining coloured thin lines correspond to simulated pits. remaining coloured thin lines correspond to simulated pits.
M GODOY
Recovered Gold
350
300
250
Ore (Mt)
200
ResSim1 ResSim2
150 ResSim3 ResSim4
ResSim5 ResSim6
100 ResSim7 ResSim8
ResSim9 ResSim10
ResSim11 ResSim12
ResSim15 EstMod
0
1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 45 47 49 51
Pit Number
FIG 9 - Total recovered metal contained on incremental pit shells.

Estimated pit is depicted by the thick orange line and the
remaining coloured thin lines correspond to simulated pits.
Undiscounted Cash Flow (Net Value)

1400
1200
1000
Net Value (M$)
800
ResSim1 ResSim2
600
ResSim3 ResSim4
ResSim5 ResSim6
400 ResSim7 ResSim8
ResSim9 ResSim10
ResSim11 ResSim12
ResSim15 EstMod
0
1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 45 47 49 51
Pit Number
FIG 10 - Total pit value of incremental pit shells. Estimated pit is

depicted by the thick orange line and the remaining coloured thin
lines correspond to simulated pits.
against 9.84 MOz in the estimated pit, which is an increase of

approximately 0.9 per cent. This combination of slightly higher
metal with considerable more rock tonnage in the simulated pits
translates into a reduced net pit value when compared to the
estimated pit. The estimated pit presents a consistently higher
Net Value until pit shell 39 when the total rock tonnage becomes
equivalent to that shown on the simulated pits. For pit shell 27
the average Net Value over simulated pits is 1159 million dollars
against 1288 million in the estimated pit (approximately ten per
cent less). The risk profile on Net Value goes from $1068 M to
$1292 M, which corresponds to a range of -7.86 per cent to FIG 11 - Cross-sections produced for pit shell 27 over all
+6 per cent in relation to the expected value. It is interesting to
optimised models overlaid with actual pit design.
note that as for the estimated model the maximum Net Value
over all simulated pit corresponds to pit shell number 26. This
shows that pit shell 26 is quite robust with regards to grade extension represents a major downside potential zone and goes
uncertainty. Another conclusion that can be drawn from these from the actual pit surface to the lowest levels of the pit. The
results, which is coherent with the results obtained in the analysis simulated pits indicate an upside potential region at the
presented in the previous section, is that there is a global southwest zone of the pit where the simulated pits reach levels
overestimation of grades and ore tonnage by the estimated model that are deeper than the current pit design. The spread of cloud of
as compared to the simulated models. simulated pits is shown to be low from the surface down to level
A series of cross-sections were produced for pit shell 27 over -350. Bellow that the spread increases considerably. This reflects
all optimised models and for a pit design. These cross-sections the increasing uncertainty on the distribution of grades at depth
were overlaid and are presented in Figure 11. The main and is directed related to the lack of drilling.
conclusions drawn from the analysis of these cross-sections are
the following. The simulated pits closely honour the eastern wall TRADE-OFF ANALYSIS
of the current pit design, showing that the eastern slope is stable
with relation to grade uncertainty. In general, there is a more The goal of this analysis is to quantify the impact of grade
pronounced fluctuation in the western wall which indicates uncertainty to tonnage, grades, metal and Net Value of two
higher levels of grade uncertainty. The current design has an different mining strategies. Scenario ‘A’ considers the depletion
extension of the western wall, which is not included in of a cut back as a single stage, while scenario ‘B’ defines two
the optimisations of both estimated and simulated pits. This separate stages. The analysis consists on the quantification of
uncertainty on key mining physicals and economic parameters

for the two mining scenarios considered. The objective is to
300
evaluate if one of the scenarios is any better in terms of the
compromise between Net Value and risk exposure. The Maximum
Ore Quantity (Mt)

a
250
Minimum
procedure proposed to develop the analysis is the following:
Expected
1. generate a wireframe describing the cut-back, 200 Estimated
2. filter the block model against the wireframe and retain the 150
blocks lying inside the cut-back as new block model,
100
3. export the block model produced above into a Whittle
Model File,
50
4. repeat steps 2 and 3 for each one of the simulated models, OneStage FirstStage SecondStage
5. for each model produced in the previous steps calculate the

relevant summary information, and FIG 13 - Risk profiles on contained ore of two mining scenarios.
6. repeat steps 1 to 6 for stages 1 and 2 that correspond to

another mining scenario.
The procedure was carried out for a total of 16 models,
corresponding to 15 grade simulations plus the estimated model. 1.8 Maximum
Metal Quantity (Moz) a

Minimum
Figure 12 presents a 3D view of the cut-back showing its 1.5
Expected
position in relation to the current pit design. Figure 13 shows the 1.3 Estimated
results in terms of contained ore for each mining scenario:
1.0
• the first profile corresponds to the scenario ‘A’, which 0.8
corresponds to the cut-back depletions as a single stage; and
0.5
• the other profiles correspond to the first and second stages of
0.3
scenario ‘B’.
0.0
OneStage FirstStage SecondStage
FIG 14 - Risk profiles on recovered metal of two mining scenarios.
cent and 3.61 per cent respectively. Figure 15 summarises the

results in terms of Net Value.
The expected Net Value of scenario ‘B’ is approximately 22.3 per
cent lower than in scenario ‘A’ ($76.5 M against $98.4 M). The
risk profile for scenario ‘A’ shows a range of variation that
corresponds to 50 per cent of the expected Net Value. For
scenario ‘B’ the ranges correspond to 68.61 per cent and
141.2 per cent for the first and second stages respectively. The
combined expected range of variation for scenario ‘B’
corresponds to 67.4 per cent. For scenario ‘A’ the Net Value
predicted by the analysis on the estimated model is 76.85 per
FIG 12 - View of the cut-back aganst the pit design. The colours cent higher than the expected tonnage derived from the
indicate different incremental pit shells. simulations. For the first and second stages this difference
corresponds to 162.47 per cent and -2.48 per cent respectively.
The expected combined ore tonnage of scenario ‘B’ is 2.3 per The results obtained in terms of the risk profiles indicate that
cent higher than scenario ‘A’ (28.2 Mt against 27.6 Mt). The risk both scenarios present high risk of not achieving predicted Net
profile for scenario ‘A’ shows a range of variation that Value. In addition, it was clearly identified that the volume
corresponds to 5.2 per cent of the expected ore tonnage. For the related to stage 1 of the scenario ‘B’ is the one with all the risk.
two stages of scenario ‘B’, these ranges correspond to 5.19 per
cent and 6.51 per cent respectively. For scenario ‘A’, the
contained ore predicted by the estimated model is 4.16 per cent 200
higher than the expected tonnage derived from the simulations. Maximum
175
For the first and second stages, this difference corresponds to Minimum
Net Value ($M) a
150
4.63 per cent and 9.89 per cent respectively. Figure 14 shows the Expected
Estimated
results in terms of recovered gold. 125
The expected recovered metal of scenario ‘B’ is approximately 100
1.7 per cent higher than in scenario ‘A’ (1.63 Moz against 75
1.60 Mt). The risk profile for scenario ‘A’ shows a range of 50
variation that corresponds to 7.1 per cent of the expected 25
tonnage. For scenario ‘B’ this ranges correspond to 8.5 per cent
0
and 12.3 per cent. For scenario ‘A’ the recovered gold predicted
OneStage FirstStage SecondStage
by the estimated model is 10.43 per cent higher than the expected
tonnage derived from the simulations. For the first and second
stages of scenario ‘B’, this difference corresponds to 16.70 per FIG 15 - Risk profiles on pit value of two mining scenarios.
M GODOY
It is important to note that this analysis is rather simplistic in the value for the Net Value. The results indicate a relatively low
sense that the time effect of money is not included. Ideally, a uncertainty in tonnages and grades contained between pits 26
mining schedule should be developed in order to account for the and 27. However, this relatively low uncertainty becomes a
mine sequencing. However, the second scenario roughly critical issue due the high stripping ratio which makes the
accounts for the sequencing by developing the depletion in two increment’s Net Value very sensitive to grade uncertainty as well
stages. The analysis indicated that the risk of missing the target as gold price.
when mining the volume related to the first stage of the second Several cross-sections have been generated to show the pit
scenario is extremely high. This warrants a detailed review of the region relative to the increment between pits 26 and 27. These
estimated grades in this volume. sections show the incremental shells, the current pit design and
ore blocks contained inside the increment. The main conclusions
drawn from the analysis of these cross-sections have four
RISK ANALYSIS ON ORE BLOCKS DRIVING components. First, the major difference between the incremental
A PIT INCREMENT Whittle pit shells $550/oz and $560/oz corresponds to a region
located at the southern end of the pit. Most of the incremental ore
The aim of this analysis is the quantification of uncertainty on blocks have an expected value inside the range of 1.5 g/t to
the ore blocks driving the increment between two successive 2.5 g/t. Moreover, most of the incremental ore blocks have more
Whittle pit shells. To assess uncertainty on the main parameters than 60 per cent chance of being above the cut-off. The
driving increment between two successive pit shells, the follow increment contains a high quantity of waste and the pit design
procedure is proposed: has considerably more waste than pit shell 27.
1. From the Whittle result file produced by the pit
optimisation process generate a Whittle pit list file CONCLUSIONS
containing information about the smallest numbered pit that The goal of this work was to illustrate different applications of
each block is part of. risk analysis on the effects of grade uncertainty to various
2. Use the re-blocking program – apply the pit list file aspects of pit optimisation and design. Four cases have been
produced in step 1 to each one of the simulated block carried out to illustrate different types of applications:
models. This will create a set of results files.
1. The first case consisted of an uncertainty analysis on pit
3. Produce cross-sections for the incremental pit shells. optimisation results – Net Value, tonnage, grade and metal.
The procedure consisted in applying a set of incremental pit
4. Generate the summary pit information for the two pits. The
shells, as produced by the pit optimisation process, to a set of
analysis must be carried out for each one of the available
simulated resource models. The subsequent analysis on each
simulated models. Derive the information referent to the
model produced a set of equally probably outcomes for the
incremental volume by subtracting the cumulative mining
mining physicals and economic forecasts given the initial set
physicals (ore and metal quantities) and economic values
of incremental pit shells.
(Net Value, processing cost and mining cost) between the
two successive pit shells. 2. The second case identified areas were grade uncertainty has
The above procedure is similar to the procedure used in the major impact to the definition of the ultimate pit limits
first section of this paper. The difference is that here the analysis (upside/downside potential). Rather than developing a risk
analysis on a given set of incremental pit shells, this
is limited to two specific pit shells. The increment from pit 26 to
procedure consisted in the generation of alternative sets of
pit 27 contains approximately 53 Mt of rock and is located in the
incremental shells, one set for each simulated model.
southern end of the pit. Table 1 presents the results obtained by
the analysis of the pit increment. 3. The third case aimed at quantifying the impact of grade
uncertainty to tonnage, grades, metal and Net Value of two
different mining scenarios for a given cut-back. The main
TABLE 1 objective was to evaluate if one of the scenarios was any
Risk profile for mining physicals and economic parameters for better in terms of the compromise between Net Value and
the increment between pits 26 and 27. risk exposure.
Expected Minimum Maximum Range 4. The fourth case consisted of a risk analysis related to pit
Ore (×1000t) 5925 5706 6222 516 increments. The objective of this analysis is the
quantification of uncertainty on the ore blocks driving the
Metal (oz) 341 739 309 451 362 982 53 531
increment between two successive Whittle pit shells.
Grade (g/t) 1.95 1.93 1.98 0.05
The results were presented in two steps:
Mcost (×1000$) -103 594 -103 716 -103 521 195
1. first, each optimisation output was evaluated in terms of
Pcost (×1000$) -70 783 -74 285 -68 238 6047
contained ore, grade, metal and pit value; and
Net value (×1000$) 7381 -8164 17 685 25 849
2. the second step of the analysis consisted on the generation
of a series of cross-sections.
The risk profile on the contained ore shows a range between These cross-sections were taken over all optimised models and
5.7 Mt and 6.2 Mt, with an expected tonnage of approximately included the actual pit design. Several conclusions have been
5.9 Mt. Low ranges of variation are also shown for the risk drawn from these graphs indicating areas of upside and downside
profiles on grade, metal content and costs. The main issue here is potential.
in relation to the Net Value, which has a chance of being
negative. Its risk profile goes from approximately $-8.2 M to This paper presented a set of procedures that enable mine
$17.7 M. The reason for the increase in the risk profile the planning engineers to carry out a series of analysis, which can be
mining cost associated to a high stripping ratio (~9), which used to evaluate the sensitivity of incremental pit shells and pit
makes the Net Value oversensitive to possible variations on the designs to grade uncertainty. The results obtained from the
recovered gold. It is important to notice that the risk profile analysis have shown to provide valuable information, which can
indicates an expected Net Value for this increment of $7.4 M. In be used to develop mining strategies that are risk resilient in
fact, only one out of 15 simulated models presented a negative relation to grade uncertainty.
REFERENCES Kent, M, Peattie, R and Chamberlain, V, 2007. Incorporating grade

uncertainty in the decision to expand the main pit at the Navachab
David, M, 1988. Handbook of Applied Advanced Geostatistical Ore gold mine, Namibia, through the use of stochastic simulation, in
Reserve Estimation, 216 p (Elsevier Science Publishers: Amsterdam). Orebody Modelling and Strategic Mine Planning, second edition
Dimitrakopoulos, R, 1998. Conditional simulation algorithms for (ed: R Dimitrakopoulos), pp 207-216 (The Australasian Institute of
modelling orebody uncertainty in open pit optimisation, Mining and Metallurgy: Melbourne).
International Journal of Surface Mining, Reclamation and Lerchs, H and Grossmann, I F, 1965. Optimum design of open pit mines,
Environment, 12:173-179. CIM Bulletin, Canadian Institute of Mining and Metallurgy, volume
Dimitrakopoulos, R, Farrelly, C T and Godoy, M, 2002. Moving forward 58, January.
from traditional optimisation: grade uncertainty and risk effects in Ravenscroft, P J, 1992. Risk analysis for mine scheduling by conditional
open-pit design, Transactions of the Institutions of Mining and simulation, Transactions of the Institutions of Mining and
Metallurgy, Mining Technology, 111:A82-A88. Metallurgy, Mining Technology, 101:A101-A108.
Dimitrakopoulos, R, Martinez, L and Ramazan, S, 2007. Optimising open Whittle, J, 1999. A decade of open pit mine planning and optimisation –
pit design with simulated orebodies and Whittle Four-X: A The craft of turning algorithms into packages, in Proceedings
maximum upside/minimum downside approach, in Orebody APCOM ’99 International Symposium on Application of Computers
Modelling and Strategic Mine Planning, second edition (ed: R and Operations Research in the Mineral Industry), pp 15-24
Dimitrakopoulos), pp 201-206 (The Australasian Institute of Mining (Colorado School of Mines: Colorado).
and Metallurgy: Melbourne).
Godoy, M C, 2003. The effective management of geological risk in
long-term production scheduling of open pit mines, PhD thesis,
256 p, The University of Queensland, Brisbane.
Journel, A G, 1992. Computer imaging in the minerals industry – Beyond
mere aesthetics, in Proceedings 23rd APCOM (International
Symposium on Application of Computers and Operations Research in
the Mineral Industry), pp 3-13 (Society for Mining, Metallurgy and
Exploration: Littleton).
HOME
Accounting for Joint Ore Supply, Metal Price and Exchange Rate
Uncertainties in Mine Design
S A Abdel Sabour1 and R Dimitrakopoulos2
ABSTRACT Among other sources, the geological and market uncertainties

are major sources of risk in mining businesses. An example
Conventional methods for selecting a production plan for an open pit for geological uncertainty is illustrated in Figure 1, showing 50
mine are based on the assumption that ore grades throughout mine life as
simulations for the grade of an Australian gold mine. It is obvious
well as market variables are known with certainty. Another important
issue that is also being ignored in the conventional procedure is the value
that the possible average mill feed grade is highly uncertain.
of management flexibility to revise operating policy with time according Ignoring such uncertainty about ore tonnage and grade could
to the arrival of new information. Based on this oversimplification of result in substantial losses especially when considering the
reality, an optimum mine plan that is thought to be maximising net capital-intensive nature of mining investments. As reported by
present value (NPV) of the project is selected. Given the multiple sources Vallee (2000), a World Bank study conducted in the 1990s showed
of risk affecting the estimated NPV, there is a high probability that the that a total loss of $US1106 million resulted from ore reserve
actual value will be different from that estimated. This could mislead problems in failed mining projects in Canada and the USA.
decision making process regarding selecting an optimum mine plan.
This paper proposes a multi-criteria ranking system for selection 2.10
between alternative mine designs under both geological and market
uncertainties. The system is based on integrating multiple market and 2.05
geological uncertainties as well as the operating flexibility to revise the
2.00
ultimate pit limits using a Monte Carlo-based real options valuation
Mill Feed Grade, g/t Au
(ROV) model. To compare the methods, the article applies the proposed 1.95
system, along with the conventional NPV-based method, to rank possible
1.90
mine designs at a Canadian copper mine. It has been found that,
considering multiple risk analysis measures while integrating uncertainty 1.85
and operating flexibility results in improving the mine design selection
1.80
process.
1.75
INTRODUCTION 1.70
1.65
The ultimate goal in open pit mine planning is to define an
‘optimum’ mine design and a life-of-mine (LOM) sequence of 1.60
production. In this respect, the conventional method for open pit Possible Average Mill Feed Grade
planning starts with modelling the orebody based on the borehole
data and geological information. Then, the mining field is
FIG 1 - Uncertainty of ore grade (modified from Dimitrakopoulos,
divided into blocks of regular volume. Based on a deterministic
Farrelly and Godoy, 2002).
metal price, each block is assigned a value equal to the gross
value of its metal content minus the applicable production,
processing and refining costs. The optimum production plan is The second major source of risk affecting mine project
determined by applying different heuristic and optimisation profitability is related to the uncertainty about market behaviour
algorithms (Whittle, 1988). Commonly, there may be alternative of metal prices and exchange rates. Figure 2 shows, for example,
technically feasible mine plans available that meet operational the average annual market prices of gold and copper over the
and technical constraints. The selection among those plans is period 1980 - 2006 in current US dollars. It is obvious that these
then based on economic reasons. This is carried out by prices are highly volatile and do not keep a constant trend, which
evaluating each of the possible mine plans and comparing their makes it speculative to define deterministic forecasts for future
economic attractiveness so as to select the most economically metal prices. Exchange rate is also another contributor to project
risk. Figure 3 shows how the exchange rates of the currencies of
appealing one. The basic assumption in most previous work is
two major metal exporting countries like Australia and Canada
that both the grades of blocks and metal prices are known with are uncertain. Therefore, it is difficult for mine planners to have
certainty (see for example Whittle, 1988; Ramazan, 2007). precise forecasts for these key variables over the mine life.
Nevertheless, the real situation in mining industry is much more
Under the uncertain geological and market conditions, the
complex than this simple assumption suggests. In practice, mine process of selecting a mine plan among different alternatives is
planners cannot know with certainty the quantity and quality of not an easy task. Conventional static financial evaluation
ore in the ground. Also, both future metal prices and foreign methods based on deterministic geological and market variables
exchange rates are impossible to know with certainty. are not well suited to handle the multiple sources of risk and the
dynamic, proactive, nature of management decisions. There is a
need for a more efficient system for mine plan selection under
1. COSMO – Stochastic Mine Planning Laboratory, Department of multiple uncertainties that minimises the need for subjective
Mining and Materials Engineering, McGill University, FDA judgements. Such system should be based on sound
Building, 3450 University Street, Montreal QC H3A 2A7, Canada. quantification and management of both the geological and
Email: [email protected] market uncertainty. Integrating the geological uncertainty into
2. FAusIMM, Professor and Director, COSMO – Stochastic Mine open pit mine planning was first introduced by Dimitrakopoulos,
Planning Laboratory, Department of Mining and Materials Farrelly and Godoy (2002). Godoy and Dimitrakopoulos (2004)
Engineering, McGill University, Montreal QC H3A 2A7, Canada. and Leite and Dimitrakopoulos (2007) presented a risk-based
Email: [email protected] optimisation approach for long-term open pit mine planning
S A ABDEL SABOUR and R DIMITRAKOPOULOS
700 3.40 This paper builds on the work of Dimitrakopoulos, Martinez

Gold price and Ramazan (2007) by quantifying and integrating the market
Copper price 3.00 uncertainty related to metal prices and exchange rates into mine
600 planning. In this respect, the article aims to develop a system for
Copper price, US$/lb

mine plan selection based on multiple value statistics and cash
Gold price, US$/oz
2.60
flows characteristics incorporating the value of management
500
2.20 flexibility to react to the new information. In the next sections,
the proposed selection system will be briefly outlined. Then, it
1.80 will be applied, along with the conventional methods of
400
selection, to select the optimum production schedule for a
1.40
Canadian copper mine. Finally, an investigation of the usefulness
300 of the developed system will be provided.
1.00
200 0.60 A SYSTEM FOR MINE DESIGN SELECTION

1980 1985 1990 1995 2000 2005 UNDER UNCERTAINTY
Year
First, it is worth noting that the different alternative mine designs
FIG 2 - Gold and copper prices uncertainty. could result from different available production scenarios,
different LOM sequences, and so on. In this study, the procedure
described in Dimitrakopoulos, Martinez and Ramazan (2007) for
2.2 generating different mine designs based on simulating multiple
US$/AU$
orebody realisations is used. The mine design selection system,
2 US$/CAN$ or ranking system, proposed in this work takes into account
multiple sources of uncertainty simultaneously and integrates the
operating flexibility to revise the ultimate pit limits based on the
Exchange rate
1.8
new information. The system is outlined in Figure 4 and consists
of three main steps: uncertainty quantification, design valuation
1.6
and design ranking.
1.4
Quantifying geological
1.2 uncertainty
1
Jan-1990 Jan-1993 Jan-1996 Jan-1999 Jan-2002 Jan-2005 Quantifying exchange rate Quantifying metal prices
Time uncertainty uncertainty
FIG 3 - Exchange rates uncertainty.

RO valuation using Monte
considering geological uncertainty. Dimitrakopoulos, Martinez Carlo
and Ramazan (2007) proposed an approach for designing open
pit mines under geological uncertainty based on maximising
upside potential and minimising downside risk.
Design ranking
The other significant source of uncertainty that should be
considered is the market uncertainty. In this respect, the real
options valuation (ROV) can provide a promising tool for better
evaluating alternative production schedules under uncertainty. FIG 4 - Outline of the proposed mine design selection system.
The main advantage of the ROV over the conventional financial
analysis methods based on the discounted cash flow analysis lies Uncertainty quantification
in its ability to evaluate the management flexibility to react to the
new information (Samis and Poulin, 1998 and Monkhouse and In this step, both the geological and market uncertainties are
Yeates, 2005). For more details on real options valuation see, for quantified. First, the geological uncertainty is explored by
example, Brennan and Schwartz (1985), Dixit and Pindyck simulating multiple orebody realisations based on the borehole
(1994), Tufano and Moel (1999), Slade (2001), Kamrad and data using conditional simulation (see Dimitrakopoulos, 2007).
Ernst (2001), Moel and Tufano (2002), Samis et al (2003; 2006). Market uncertainty about metal prices and foreign exchange rates
It is worth mentioning here that both the conventional net present is quantified using also stochastic modelling based on historical
value (NPV) and the ROV calculate the discounted, net present market behaviour. Examples of stochastic models include
value of future cash flows. The two techniques differ in the way Geometric Brownian Motion (GBM) model in Equation 1 (Dixit
they apply discounting and in the way they deal with operating and Pindyck, 1994) and the mean-reversion model in Equation 2
flexibility. Conventional NPV applies a single discount rate to all (Schwartz, 1997):
cash flow components and does not handle flexibility. ROV
allows for differential discounting based on specific risk of each dP
cash flow component and can capture the value of operating = αdt + σdz (1)
flexibility. Throughout this paper the abbreviation ‘NPV’ will be P
used to indicate the conventional, single discount rate, no
dP
flexibility, NPV and the abbreviation ‘ROV’ will be used to = κ(μ − 1nP)dt + σdz (2)
indicate the advanced, multiple-rates, flexibility-based NPV. P
ACCOUNTING FOR JOINT ORE SUPPLY, METAL PRICE AND EXCHANGE RATE UNCERTAINTIES IN MINE DESIGN
where: could be in different forms such as the simple power series,

Laguerre polynomials or linear combinations of different forms
P is the price (Longstaff and Schwartz, 2001). The parameters of this function
α is the expected trend can be estimated at each time by least squares regression.
σ is the standard deviation This process is performed throughout the simulated paths N at
each period and the cash flows are defined based on the optimal
dz is an increment in a standard Weiner process decision reached. If the time step is set to be one year, the
dt is an increment of time outputs of the valuation model are: the annual discounted cash
flows, the probability that the mine will be producing every year,
In Equation 2, µ is the logarithm of the long-term equilibrium the expected value of positive annual cash flows with their
level of metal price and κ is the reversion speed. It is worth corresponding probabilities and the expected value of negative
noting here that the stochastic models in Equations 1 and 2 are annual cash flows with their corresponding probabilities. These
provided as examples for illustration purposes only. Other statistics are produced for each design in addition to the expected
models can replace the above ones without having effect on the overall present value for each design along with its confidence
ideas and concepts presented in this work. The advantage of the limits at a specified level of confidence. These statistics are used
simple models presented in Equations 1 and 2 over the more to rank the different designs as explained in the following
complex models is that the model parameters can be easily section.
estimated from historical data. More details about stochastic
models and parameter estimation can be found in Dixit and Ranking of mine designs
Pindyck (1994) and Schwartz (1997), among others.
The ranking procedure proposed in this study aims to gather
multiple value and risk analysis indicators into one quantitative
Valuation of mine designs measure while integrating real industry complexities such as
The economic valuation model is based on real options uncertainty and operating flexibility to revise predefined pit
valuations using the least squares Monte Carlo method. This limits. The proposed design ranking system for selecting the best
method was developed by Longstaff and Schwartz (2001) for open pit design based on the information available at the initial
valuing American-style securities and extended by Abdel Sabour planning time takes into account the following aspects:
and Poulin (2006) to valuing capital investments under multiple • upside potential that measures the ability of designs to
market uncertainties and extended further by Dimitrakopoulos capture possibly more profits than those expected if
and Abdel Sabour (2007) to valuing mining investments under outcomes were favourable;
multiple market and geological uncertainties. In this work, the • downside risk that reflects the difference between designs in
operating flexibility to modify pit limits will be integrated into minimising negative cash flows risk throughout mine life;
the value estimate of designs taking into account metal prices,
exchange rates and geological uncertainties simultaneously. • probability of completion, which is the probability that the
mine will be open throughout its planned life; and
As described above, metal prices and exchange rates
uncertainties can be quantified using the stochastic models • statistics of the estimated values which includes the average,
presented in Equations 1 and 2 while the geological uncertainty lower and upper limit at a certain confidence level.
is quantified using the conditional simulation technique. It is After estimating the above described four indicators for each
assumed that there is M different feasible mine designs and each mine design, the total ranking indicator is simply the summation
design has H different grades and tonnages equal to the number of these four indicators. The designs are then ranked according to
of generated orebody realisations. For market variables, a large the total indicator and the design with the highest indicator
number of correlated realisations, N, are generated using the should be selected.
stochastic models and taking into account the correlation In this work, the average performance of all feasible mine
coefficients between those variables. Due to practical and designs will be used to compare and rank these designs.
computational considerations the number of market realisations, However, other references can replace these averages depending
N, is much larger than the number of orebody realisations, H. the goals and objective specific to each mining project or a
Therefore, to have N possible grades and tonnages for each mining company. Also, in this article, the above mentioned four
design throughout the mine life, N independent draws are ranking criteria are given equal weights. In other applications,
generated randomly from the H simulated orebody models. The different weights can be given to those criteria based on the
decision whether to keep or revise the originally defined pit specific policy of a mining company.
limits is taken at discrete regularly-spaced time points based on
the expected continuation value, CV, such as: CASE STUDY – SELECTING A DESIGN
Keep the predefined pit limits if: FOR A CANADIAN COPPER MINE
In this section the proposed design ranking procedure will be
E (CV n, t) > 0 (3) applied to a case study of a copper mine in Canada. It is assumed
that there are alternative technically feasible mine designs
Revise the predefined pit limits if: available to decision-makers. These different mine designs could
be generated by a number of ways. One way, which is applied in
E (CV n, t) ≤ 0 (4) this study, is to use the method described in Dimitrakopoulos,
Martinez and Ramazan (2007) for generating different designs
According to Equations 3 and 4, if the expected present value from the multiple orebody simulations. These designs have to be
at time t and sample path n is positive, the optimum decision is to ranked so as to select the best one based on the information
keep the predefined pit limits until next decision time. Otherwise, available at the planning time. To investigate the efficiencies of
if the expected present value is negative, the original plan should the valuation techniques and the proposed ranking system in
be revised and the current pit limit at time t should be the final ranking alternatives under uncertainty, the mine designs will be
pit limit. Estimating the expected CV requires knowing the ranked based on the following four ranking measures:
function that relates the present value of future mining operations • the expected value estimated by the conventional NPV
beyond time t to the states prevailing at time t. This function valuation method;
• the NPV-based indicator explained above, except the TABLE 1

probability of completion indicator since the NPV of the Economic parameters for the copper mine.
mine calculated at the planning time before starting
production does not consider the flexibility to revise pit Item Description
limits in the future; Risk-free interest rate, % 9.20
• the expected value estimated by the real options valuation; Inflation, % 6.70
and Income taxes, % 40.00
• the real options valuation (ROV) -based indicator. Initial copper price, $US/lb 1.07
It is assumed that the mine was planned and mined out in the Volatility, %, copper price 20.00
past. Therefore, it is possible to compare the design ranking
based on the actual market realisations to the ranking estimated Reversion speed of copper price 0.19
by the four methods listed above. No capital expenditures are Long-term copper price, $US/lb 1.00
considered since it is assumed that production operations are Initial $US/$CAN rate 1.20
carried out by contractors.
Volatility, %, $US/$CAN 4.00
This copper mine is assumed to be planned in 1992, started
production at the beginning of 1993 and closed at the end of Reversion speed/year, $US/$CAN 0.08
2000. The economic data prevailing at the planning time (1992) Long-term, $US/$CAN 1.25
are listed in Table 1. Both of the copper price and the exchange
rate are modelled with the mean reverting process in Equation 2.
Except the initial copper price and the initial exchange rate, all with no loss offset. At the planning time, there were ten mine
parameter estimations are based the historical data from 1970 to designs estimated by feeding the ten simulated orebody models
1992. The overall income tax rate is assumed to be 40 per cent into Whittle Software that uses the nested Lerchs-Grossman
TABLE 2
Average annual tonnage, operating cost and grade for the technically feasible copper mine designs.
Y1 Y2 Y3 Y4 Y5 Y6 Y7 Y8
Design 1 Tonnage, tonne 7.40 7.50 5.93 7.50 7.50 7.46 7.50 4.89
Cost, $CAN/tonne 14.62 14.72 15.52 13.85 14.53 14.88 13.38 12.42
Copper grade, % 0.70 0.67 0.57 0.59 0.68 0.72 0.53 0.65
Cost, $CAN/tonne 14.70 14.65 14.62 13.96 14.84 15.32 12.83 12.42
Copper grade, % 0.70 0.65 0.61 0.61 0.71 0.58 0.58 0.70
Cost, $CAN/tonne 14.42 14.64 14.62 15.03 13.26 15.28 14.29 12.44
Copper grade, % 0.70 0.66 0.63 0.56 0.63 0.78 0.51 0.63
Cost, $CAN/tonne 14.33 14.57 14.72 16.10 13.81 13.12 14.83 12.59
Copper grade, % 0.68 0.66 0.67 0.51 0.58 0.68 0.71 0.57
Cost, $CAN/tonne 14.95 14.85 14.30 13.31 14.76 15.66 13.36 12.54
Copper grade, % 0.74 0.62 0.55 0.63 0.69 0.65 0.56 0.70
Cost, $CAN/tonne 14.12 14.73 15.34 13.66 14.22 14.85 13.09 12.50
Copper grade, % 0.70 0.67 0.54 0.60 0.68 0.68 0.58 0.70
Cost, $CAN/tonne 14.42 13.51 14.63 15.55 13.68 15.08 13.93 12.43
Copper grade, % 0.67 0.65 0.67 0.59 0.61 0.77 0.52 0.65
Cost, $CAN/tonne 14.90 15.09 15.88 14.01 12.93 14.29 14.57 12.46
Copper grade, % 0.73 0.67 0.50 0.58 0.64 0.72 0.62 0.62
Cost, $CAN/tonne 14.76 14.65 14.89 14.18 12.76 15.48 14.29 12.47
Copper grade, % 0.69 0.64 0.62 0.56 0.67 0.78 0.52 0.65
Cost, $CAN/tonne 14.71 13.58 14.60 15.00 13.89 14.58 14.12 12.40
Copper grade, % 0.68 0.65 0.67 0.61 0.60 0.75 0.53 0.65
ACCOUNTING FOR JOINT ORE SUPPLY, METAL PRICE AND EXCHANGE RATE UNCERTAINTIES IN MINE DESIGN
algorithm to generate the optimum mine design (Whittle, 1999). 120

Table 2 lists the average annual tonnage, operating cost, in real
terms, and copper grades of the multiple simulated orebody 100
ROV
models for the ten designs. 80
Conventional NPV
Total ranking indicator

Each of the ten designs has been evaluated using Monte Carlo 60
method with 20 000 correlated simulations for the copper prices
and the exchange rates throughout the eight-year mine life. To 40
integrate geological uncertainty, 20 000 random independent 20
draws have been drawn from the ten simulated orebody
0
models. Therefore, there are 20 000 different possible scenarios
1 2 3 4 5 6 7 8 9 10
throughout the mine life with each scenario has its own copper -20
prices and exchange rates that are simulated using the stochastic
-40
models and an independent draw from the ten simulated ore
tonnages, production costs and ore grades. Two valuation -60
techniques are applied. The first technique is the conventional -80
NPV method that does not take into account operating flexibility Design
to revise pit limits. The second one is the ROV in which the
flexibility to revise pit limits is integrated and the optimum
decision is made based on the expected value conditional on time FIG 6 - Design indicators based on the NPV and the real
and the simulated paths. To focus on the value of operating options results.
flexibility only, the same discounting procedure was applied for
both the conventional NPV and the ROV methods. Therefore, the
difference between the two techniques here is owed solely to the TABLE 3
difference in the way they handle flexibility in pit design. Designs ranking for the copper mine.
Figure 5 shows the expected design values estimated by the
conventional static NPV and the ROV. Based on the NPV, Rank NPV NPV-based ROV ROV-based Actual
Design 6 has the highest value while the ROV indicates that expected TRI expected TRI data
Design 10 has the highest value. value value ranking
1 Design 6 Design 6 Design 10 Design 10 Design 2
45
ROV 2 Design 10 Design 10 Design 7 Design 2 Design 10
40 Conventional NPV
35 4 Design 5 Design 5 Design 6 Design 3 Design 4
Value, $CAN million

20
15
10
5
0 while it is ranked the second based on the actual market data.

1 2 3 4 5 6 7 8 9 10
Comparing the rank for all other designs, it is obvious that design
Design
ranking based on ROV is closer to the ranking based on actual
market data. Another important conclusion can be drawn from
FIG 5 - Designs valuation with the NPV and the real options. Table 3 is that, for the conventional NPV analysis, design ranking
based on the NPV expected value is almost identical to that of the
NPV-based indicator and differs from those of the real options
The advantage of Monte Carlo method is that, it provides expected value and the ROV-based indicator. The ROV expected
multiple statistics for the overall value as well as annual cash value and the ROV-based indicator have similar ranking for five
flows that can be used to carry out advanced risk analysis. These mine designs and different ranking for the remaining five designs.
statistics were used to construct a ranking system. Following the
The results show that under the conditions of uncertainty, design
proposed design ranking procedure outlined above, a NPV-based values based on actual market data can be significantly different
indicator and a ROV-based indicator have been calculated for from those estimated at the planning time. Consequently, using the
each design. Figure 6 depicts the calculated Total Ranking expected value to rank possible mine designs may result is
Indicators (TRI) based on the valuation results of NPV and ROV. suboptimal decisions. Therefore, it is important to integrate
It is worth noting that the TRI for each design is the summation multiple risk and cash flow analysis into mine design selection
of the subindicators related to the upside potential, downside process. However, given the multiple sources of uncertainty
risk, probability of completion and value statistics. The only associated with mining investments, the multiple risk and cash
difference between the NPV and the ROA indicators is that the flow analysis could be useless if the management flexibility to
NPV one consists of three subindicators since the probability of react to the new information is not considered. As shown in
completion indicator is the same for all designs. Table 3, the NPV-based indicator using multiple risk and cash flow
With the valuation results in Figure 5 and the indicators in analysis generated almost the same decision as the NPV expected
Figure 6, it is possible to rank the ten mine designs based on the value. This is because if no flexibility to revise pit limits is
four different measures, as shown in Table 3. The last column lists considered, the distribution of any estimated measure may
design ranking based on the actual market data over the mine life assumed to be approximately symmetrical around the expected
(Y 1992 - Y 2000). As indicated in Table 3, Design 6 is ranked the value and no significant change in the decision nor any
first based on NPV, while based on the actual market data it is improvement in the selection process might be achieved. On the
ranked the seventh. Based on ROV, Design 10 is ranked the first contrary, when integrating one kind of operating flexibility, which
is the flexibility to revise the originally taken decisions regarding Dimitrakopoulos, R and Abdel Sabour, S A, 2007. Evaluating mine plans
the ultimate pit limits, the efficiency of the selection process was under uncertainty: Can the real options make a difference?,
significantly improved when considering multiple risk and cash Resources Policy, 32:116-125.
flow analysis measures. Effort has to be devoted to integrate more Dimitrakopoulos, R, Farrelly, C T and Godoy, M, 2002. Moving forward
risk analysis measures as well as more operating flexibilities, such from traditional optimisation: Grade uncertainty and risk effects in
as the flexibilities to expend, contract, switch cut-off grades and open pit design, Transactions of the Institutions of Mining and
switch production inputs in order to improve further the process of Metallurgy, Mining Technology, 111:A82-A88.
mine planning under uncertainty. Dimitrakopoulos, R, Martinez, L S and Ramazan S, 2007. A maximum
upside/minimum downside approach to the traditional optimisation
The results of this study are specific to the presented case of open pit mine design, Journal of Mining Science, 43:73-82.
study and the data fed into the valuation model. Different results Dixit, A K and Pindyck, R S, 1994. Investment Under Uncertainty
might be obtained if the mine was assumed to be started and (Princeton University Press: New Jersey).
finished at different calendar times. Also, the valuation results Godoy, M C and Dimitrakopoulos, R, 2004. Managing risk and waste
are sensitive to the chosen stochastic model for describing the mining in long-term production scheduling, SME Transactions,
evolution of a market variable and its estimated parameters. In 316:43-50.
this respect, it is worth stressing that the ROV is a valuation Kamrad, B and Ernst, R, 2001. An economic model for evaluating mining
rather than a price forecasting technique. Like the static NPV, or and manufacturing ventures with output yield uncertainty, Oper Res,
even any other system, the outputs are dependent on the inputs. 49(5):690-699.
Leite, A and Dimitrakopoulos, R, 2007. A stochastic optimisation model
CONCLUSIONS deposit, Transactions of the Institutions of Mining and Metallurgy,
Mining Technology, 116:A109-A118.
A ranking system was outlined for selecting open pit mine designs
using multiple risk and cash flow analysis measures while Longstaff, F A and Schwartz, E S, 2001. Valuing American options by
simulation: A simple least-squares approach, The Review of
integrating market and geological uncertainty and the operating Financial Studies, 14:113-147.
flexibility to revise pit limits. The proposed system is based on
Moel, A and Tufano, P, 2002. When are real options exercised? An
valuing the possible, technically feasible, mine designs using a empirical study of mine closings, Review Financial Studies,
simulation-based real options valuation method that can integrate 15(1):35-64.
multiple sources of uncertainty related to market and geological Monkhouse, P H L and Yeates, G, 2007. Beyond naive optimisation, in
variables as well as the management responses to the new Orebody Modelling and Strategic Mine Planning, second edition (ed:
information in the future. This economic valuation model R Dimitrakopoulos), pp 3-8 (The Australasian Institute of Mining and
generates multiple statistics for cash flows and operating modes of Metallurgy: Melbourne).
each design throughout the active life of the project. Such statistics Ramazan, S, 2007. The new fundamental tree algorithm for production
are used to construct a system for design ranking under scheduling of open pit mines, European Journal of Operations
uncertainty. Research, 177 1153-1166.
For the sake of comparison, the proposed system was applied Samis, M, Davis, G A, Laughton, D and Poulin, R, 2006. Valuing
along with other three ranking methods, based on the NPV uncertain asset cash flows when there are no options: A real options
expected value, the ROV expected value and the NPV-based approach, Resources Policy, 30:285-298.
indicator, to rank possible designs at a copper mine. The results Samis, M, Laughton, D and Poulin, R, 2003. Risk discounting: The
show that the design selection process may improve when fundamental difference between the real option and discounted cash
incorporating uncertainty and operating flexibility. Future flow project valuation methods [online]. Available from: <http://
papers.ssrn.com> [Accessed: 18 November, 2008].
extensions could include integrating more risk analysis measures
Samis, M and Poulin, R, 1998. Valuing management flexibility: A basis
and other types of management flexibilities such as the flexibility
to compare the standard DCF and MAP frameworks, CIM Bulletin,
to revise the cut-off grade with time so as to improve the design 91:69-74.
selection process even further. Schwartz, E S, 1997. The stochastic behaviour of commodity prices:
Implications for valuation and hedging, Journal of Finance,
ACKNOWLEDGEMENTS 52:923-973.
Slade, M E, 2001. Valuing managerial flexibility: an application of real
The work in this paper was funded from NSERC CDR Grant option theory to mining investments, Journal of Environmental
335696 and BHP Billiton, as well NSERC Discovery Grant Economics and Management, 41:193-233.
239019, McGill’s COSMO Lab and its industry members, Tufano, P and Moel, A, 1999. Bidding for Antamina: Incentives in a real
AngloGold Ashanti, Barrick, BHP Billiton, De Beers, Newmont, option context, in Project Flexibility, Agency and Competition, (eds:
Vale and Vale Inco. M J Brennan and L Trigeorgis), pp 128-150 (Oxford University
Press: Oxford).
Vallee, M, 2000. Mineral resource + engineering, economic and legal
REFERENCES feasibility = ore reserve, CIM Bulletin, 93:53-61.
Abdel Sabour, S A and Poulin, R, 2006. Valuing real capital investments Whittle, J, 1988. Beyond optimisation in open pit design, in Proceedings
using the least-squares Monte Carlo method, The Engineering Canadian Conference on Computer Applications in the Mineral
Economist, 51:141-160. Industries, pp 331-337 (Balkema: Rotterdam).
Brennan, M J and Schwartz, E S, 1985. Evaluating natural resource Whittle, J, 1999. A decade of open pit mine planning and optimisation –
investments, Journal of Business, 58:135-157. The craft of turning algorithms into packages, in Proceedings
Dimitrakopoulos, R, 2007. Applied risk assessment for ore reserves and APCOM ’99 – Computer Applications in the Minerals Industries
mine planning, Professional development short course notes, 350 p 28th International Symposium, pp 15-24 (Colorado School of Mines:
(The Australasian Institute of Mining and Metallurgy: Melbourne). Golden).
HOME
Production Scheduling Under Metal Uncertainty – Application of

Stochastic Mathematical Programming at an Open Pit Copper
Mine and Comparison to Conventional Scheduling
A Leite1 and R Dimitrakopoulos2
ABSTRACT by considering the one with a minimum downside or maximum

Life-of-Mine (LOM) production scheduling of open pit mines is of major
upside potential. Dimitrakopoulos and Ramazan (2004) present a
importance as it defines the ore supply that the mine can provide while mathematical programming formulation minimising deviations
maximising its present value. The conventional mine scheduling approach form production targets by considering a probabilistic approach.
requires a single, average type representation of reality and, as a result, it The new concept of geological discounted rate is introduced and
ignores uncertainty about technical and economic parameters. From all produces a decreasing unit cost for deviation of target over the
sources of uncertainty affecting the planned ore supply of a given LOM. Godoy and Dimitrakopoulos (2004) and Leite and
schedule, uncertainty about ore reserves is acknowledged to be the most Dimitrakopoulos (2007) propose a different stochastic approach
critical one. It is known that the use of a single, average type of deposit using a simulated annealing algorithm to obtain a stochastic
model to define reserves produces a smoothed image of the real deposit
and this smoothing may lead to a misleading forecast of ore supply, waste schedule. The proposed solution is divided into three stages:
and related cash flow over the life of the mine. To overcome this problem, 1. optimum mining rates are defined using an LP formulation,
a stochastic mining production scheduling approach is considered here.
The approach explicitly integrates the uncertainty of parameters or inputs 2. rates are utilised to schedule a set of simulated orebodies,
to the scheduling problem aimed at generating a more risk-robust and
solution. In this study, a Stochastic Integer Programming (SIP)
formulation for mining production scheduling is applied and tested at a 3. schedules are then used in a final stage in which a
low-grade variability copper deposit. The stochastic solution as stochastic schedule is obtained by using a simulated
implemented herein aims to maximise the economic value of a project annealing algorithm and the schedules as an input.
and minimise deviations from production targets in the presence of The studies show a 25 per cent to 28 per cent difference in
geological uncertainty. Unlike the conventional approach, the method is NPV as compared to the conventional schedule. Menabde et al
able to account for and manage risk.
(2007) implement a stochastic integer programming (SIP)
In the case study presented herein, the stochastic formulation for mine formulation for long-term production schedule that maximises
production scheduling is shown to produce 29 per cent higher Net Present
Value (NPV) than the schedule obtained from a conventional scheduler.
NPV considering several possible simulated orebodies and
This is an important difference, particularly under the current global simultaneously optimising cut-off grades. Ramazan and
financial climate. Dimitrakopoulos (2007a) propose an approach that accounts for
all available realisations of the orebody simultaneously in an SIP
formulation. Their formulation has as objectives the maxim-
INTRODUCTION isation of NPV and minimisation of deviation from production
The optimum open pit mine production schedule is defined as the targets. Different penalties may be defined for deviations of
sequence of extraction that maximises the present value of the different targets. Jewbali (2006) combines the use of the SIP
project. This task is one of the most challenging and important in formulation with the use of simulated future data, updating
the mine planning framework as it defines the ore supply simulated models to produce an optimum stochastic mine
produced over the LOM and consequentially has a substantial schedule and shows a 30 per cent increase in NPV at a gold
impact on the NPV of the project. The conventional mine design deposit.
and production framework defines the extraction sequence Despite substantial monetary benefits, the application of
considering a single, average type of orebody model as input. As stochastic schedulers is relatively recent and the value of this
a result, it does not account for uncertainty in the related project solution when applied to different types of deposits is still not
parameters and leads to misleading assessments (Ravenscroft, completely understood. In the present study, the approach
1992; Dimitrakopoulos, Farrelly and Godoy, 2002). This finding presented by Ramazan and Dimitrakopoulos (2007b) is applied
makes clear that there is a need to address the mine production to a low-grade variability copper deposit. The study tests the
schedule problem using stochastic approaches, which use as approach, quantifies the associated value of the stochastic
input a set of equally probable representations of the orebody solution, assesses the risk profile of pertinent mining parameters
reproducing, by construction, its actual spatial variability and and, finally, analyses the results to propose future improvements.
distribution. This set of representations is able to directly The following sections outline the stochastic integer
incorporate uncertainty into the formulation of the problem. programming formulation, present its application in a copper
Different stochastic approaches have been considered in order deposit, and compare the results with those obtained by a
to provide more robust solutions dealing with uncertainty. conventional scheduler. Conclusions follow.
Dimitrakopoulos, Martinez and Ramazan (2007) utilise the
concepts of upside/downside potential to include grade uncertainty A Stochastic Integer Programming formulation
in pushback design. Several pit designs for a set of simulated
Stochastic mathematical programming approaches to the mine
orebodies are obtained and a final pushback configuration selected
schedule problem are avenues to directly incorporate uncertainty
about ore supply in the formulation of the problem, so as to
1. Vale Ltd, Canada. Email: [email protected] minimise risk of not meeting the mine production targets. In the
mine production schedule case, the decision to be made is the
2. FAusIMM, Professor and Director, COSMO – Stochastic Mine
Planning Laboratory, Department of Mining and Materials time period in which each block is mined, in order to maximise
Engineering, McGill University, Montréal QC H3A 2A7, Canada. the overall discounted value of the project, subject to slope,
Email: [email protected] reserves and processing and mining capacity constraints. The set
A LEITE and R DIMITRAKOPOULOS
of blocks available to be scheduled are the ones contained within production target deviations. It can also be used to blend risk by
the ultimate pit. The SIP formulation presented herein includes using a low geologic risk discount rate.
uncertainty in the formulation of the problem by considering a The unit costs in Equation 1 are for excess production:
set of different and equally probable stochastic simulated
orebody realisations in the optimisation process. cu0 o
cuto = (2)
(1 + GRD)t
Objective function
The SIP formulation below aims to maximise the NPV of the where:
mine by minimising the risk of falling short of previously defined d is the discount factor mentioned earlier, or for deficient
production targets. It includes two possible destinations for a production
block:
1. processing plant, or cl0 o
clto = (3)
2. waste dump. (1 + GRD)t
The objective function includes two components and it is:
Equation 1 is subject to constraints detailed in Ramazan and
Dimitrakopoulos (2007b).
⎡ ⎤
p ⎢N m ⎥
Max ∑ ⎢∑ E {( NPV )i} * bi − ∑ ( cu d su + cl d sl ⎥ APPLICATION AT A COPPER DEPOSIT
t t to to to to
(1)
⎢1 144244 3⎥
t=1 i=l
442443 s=1
Part B
⎢⎣ Part A ⎥⎦ The deposit
The deposit is located in a typical archean greenstone belt. The
where:
region consists predominantly of mafic lavas with lesser amounts
i is the block identifier of intermediate to felsics volcaniclastics. The geological database
t is the time period; to flags the ore production target type consists of 185 drill holes with 10 m copper composites in a
pseudo-regular grid of 50 m × 50 m covering an approximately
l stands for lower bound rectangular area of 1600 × 900 m2. Using the geological inform-
u stands for upper bound ation available, one mineralisation domain is defined and modeled
through a geostatistical study. An estimated model, obtained using
s stands for the simulation number ordinary kriging, and a set of simulated models, generated using a
p is the maximum number of scheduling periods direct block simulation method (Godoy, 2003) are produced. All
models are constrained by the ore domain previously defined.
N is the total number of blocks to be scheduled Figure 1 brings a section of the estimated model and three of the
bit is a variable representing the portion of block i to be mined simulated ones. The smoothing effect on the grade, produced by
in period t the ordinary kriging model, is made clear in Figure 1.
If defined as a binary variable, it is equal to one if the block i
is to be mined in period t and equal to 0. Otherwise;
E {( NPV )ti} is the expected NPV to be generated by mining
block i in period t – it is computed as the
discounted value of Equation 1
cuto is the unit cost for excess of ore production
to
d su is the excess amount of ore production in period t
considering simulation s
clto is the unit cost for the deficient ore production
to
d sl is the deficient amount of ore production in
period t considering simulation s
The first component (Part A) in Equation 1 contributes to the
maximisation of NPV of the project. The expected NPV of a block
is computed as the expected present value if the block is mined in
period t, considering all simulated values. The second part (Part B) FIG 1 - East-West sections of three simulated models (left column)
is responsible for minimising deviations from ore production and the corresponding single estimated model (right).
targets, also managing the distribution of risk within and between
periods over the LOM. Risk management is accomplished by the
use of a geological discount rate (GDR), which discounts over Production scheduling
time the penalties applied to the unit cost deviations as explained The SIP formulation in Equation 1 used to generate the mine
below. The initial penalties for excess, cu0 o , or shortage production, production schedule first requires the definition of the ultimate
cl0 o , are user-defined and should be at the same order of magnitude pit and pushbacks. The nested pit implementation of the
as the first part of the objective function to ensure the second part Lerchs-Grossman algorithm (Whittle, 1988) for pit optimisation
is being properly considered. The impact of discounted penalties is is used here to define a final pit and pushbacks. This algorithm
a progressive decrease in the unit cost over the periods. This setup requires as an input a single representation of the orebody, which
can be used to ensure that less ‘risky’ mining blocks will be is created using ordinary kriging on a 20 × 20 × 10 m3 block
scheduled in early periods, therefore decreasing the risk of not support. The associated economic and technical parameters are
attaining the planned targets and guarantying the minimisation of given in Table 1. This model and parameters are utilised in the
PRODUCTION SCHEDULING UNDER METAL UNCERTAINTY
following sections for comparison with the results obtained by additional schedules are generated, as one of the objectives of the
the SIP. work is to test and better understand the effect of the geological
Using the parameters in Table 1 and the conventionally discount rate on the schedule produced.
estimated orebody model, a set of nested pits is generated. Pit 16
is selected as the ultimate pit limit as it corresponds to the
maximum net present value. There are 14 480 blocks inside the pit
limits. In this study, a previously established ore processing
capacity of 7.5 M tonnes per year is used. The yearly maximum
mining capacity is set to 28 M tonnes although there is no
constraint to guarantee a constant material movement over the
LOM. With this setup the scheduler is free to define the optimum
waste production strategy. Twenty realisations of the deposit are
available and generated with the direct block simulation method
(Godoy, 2003). To obtain a simulated model, the orebody is
divided into blocks of 20 × 20 × 10 m3 within the mineralised
domain. Each block is then represented by 10 × 10 × 1 nodes. This
number of nodes, 100 per block, is large enough to ensure that the
actual block scale variability is reproduced by the simulated
orebodies. These stochastic simulated orebodies are used as an
input for the SIP models and to produce risk profiles of
performance parameters of the schedules generate throughout the
study.
FIG 2 - Risk profile for ore production of the base case schedule.
TABLE 1
Economic and technical parameters.
Copper price (US$/lb) 2.0
Selling cost (US$/lb) 0.3
Mining cost ($/tonne) 1.0
Processing cost ($/tonne) 9.0
Slope angle 45°
Processing recovery 0.9
A base case SIP schedule is first considered and then the

sensitivity of pertinent key project parameters is tested and their
impact on the project’s NPV and performance evaluated. The
stochastic schedule is generated considering a 20 per cent
geological discount rate. All cases utilise the same parameters as
specified in Table 2. A fixed copper grade cut-off of 0.3 per cent is
used in combination with the probability cut-off to classify ore and
waste blocks. Penalties for excess and shortage production are
select accordingly to the magnitude of the first part of the
objective function in order to ensure that the second component, FIG 3 - Risk profile for waste production of the base case
which accounts for minimisation of production targets and risk schedule.
management, is properly weighted in the objective function. Past
work has shown that it is the order of magnitude of the penalties
rather than the actual values that effect the optimisation process. A
higher penalty for shortage production is imposed as it has a more
severe impact on the project than an excess production, which
could be handled by the use of stockpiles. Risk profiles for ore and
waste production, cumulative NPV and production deviations are
generated and presented respectively by Figures 2, 3, 4 and 5. Two
TABLE 2
SIP related parameters.
Geologic risk discount rate 20%

Cost of shortage in ore production 10 000
(unit/tonne)
Cost of excess ore production (unit/tonne) 1000
Economic discount rate (%) 10
Cut-off (% Cu) 0.3
Number of simulated orebody models 20 FIG 4 - Risk profile for cumulative NPV of the base case schedule.
A LEITE and R DIMITRAKOPOULOS
related risk profile that the conventional schedule generated

misleading results in the presence of grade uncertainty – the
forecasted ore supply has about a five per cent chance (one in 20
chances that equally likely scenarios of the deposit will produce
what is expected) for almost all years. Only in two years is the
expected forecasted production realised, year 5 and year 7. The
economic implication of such deviations can be seen in Figure 7,
where the risk profile of cumulative NPV for both the base case
stochastic schedule and the conventional schedule are shown.
The average NPV difference between the conventional and the
stochastic solution is approximately 29 per cent. The higher NPV
obtained by the stochastic solution is obtained by first
incorporating grade uncertainty into the mine production
scheduling formulation, minimising the possible deviations of ore
production target and at the same time managing the risk between
mine periods. The difference between the value of the stochastic
solution and the conventional one comes from the capacity of the
stochastic optimiser to obtain an optimum schedule considering
simultaneously several equally probable orebody models. The
FIG 5 - Deviations from ore production targets for the base case equally probable orebody models, obtained by conditional
schedule. simulation (better than the smoothed model utilised by the
conventional scheduler), represent the spatial grade distribution
of the deposit.
COMPARISON TO A CONVENTIONAL
SCHEDULER
To assess the value of the stochastic solution in the case study
presented above, a conventional LOM production schedule is
developed in this section and the economic difference between
the stochastic and conventional approaches are evaluated. The
conventional schedule is generated using the average type of
deposit model previously described and obtained using the
technique of ordinary kriging. The same final pit and push-backs
are utilised with the associated technical and economic
parameters as before (Table 1). The schedule is obtained using
Millawa NPV algorithm (Whittle, 1999) using a fixed 7.5 M
tonnes ore production target and a maximum material movement
of 28 M tonnes. No constraint or limitation is imposed on the
scheduler, so as to keep the total material movement constant
over time. The intent is to let the optimiser decide the best waste
removal strategy and to replicate the condition imposed on the
SIP scheduler.
Figure 6 presents the ore production of the conventional FIG 6 - Ore production and associated risk profile for the
schedule and its risk profile. It is clear in the figure and the conventional schedule.
FIG 7 - Conventional and SIP base case schedules and their respective risk profiles.
PRODUCTION SCHEDULING UNDER METAL UNCERTAINTY
The conventional scheduler forecasts a LOM of eight years in Dimitrakopoulos, R, Martinez, L S and Ramazan, S, 2007. A maximum
contrast with the seven years LOM forecasted by the stochastic upside/minimum downside approach to the traditional optimisation
of open pit mine design, Journal of Mining Science, 43:73-82.
scheduler. By using a smoothed scenario, the conventional
scheduler overestimates the amount of ore above the 0.3 % Cu Dimitrakopoulos, R and Ramazan, S, 2004. Uncertainty based production
scheduling in open pit mining, SME Transactions, 316:106-112.
cut-off, therefore the one extra year of production is required to
Godoy, M, 2003. The efficient management of geological risk in
fully mine the deposit. It is interesting to emphasise that this fact long-term production scheduling of open pit mines, PhD thesis,
cannot be generalised as it is a function of the grade distribution University of Queensland, Brisbane.
and cut-off applied. Godoy, M and Dimitrakopoulos, R, 2004. Managing risk and waste
mining in long-term production scheduling, SME Transactions,
CONCLUSIONS 316:43-50.
Jewbali, A, 2006. Modelling geological uncertainty for stochastic
This study shows a stochastic mine scheduling formulation and its short-term production scheduling in open pit mines, PhD thesis,
application to scheduling under grade uncertainty. The results find University of Queensland, Brisbane.
the value of the stochastic solution by comparing the results Leite, A and Dimitrakopoulos, R, 2007. A stochastic optimisation model
obtained by the stochastic schedule with the one obtained by a for open pit mine planning: Application and risk analysis at a copper
conventional scheduler. The difference is significant (29 per cent) deposit, Transactions of the Institutions of Mining and Metallurgy,
and derived from the inability of the conventional scheduler to Mining Technology, 116:A109-A118.
include grade uncertainty in its formulation, in comparison with Menabde, M, Froyland, G, Stone, P and Yeates, G, 2007. Mining
the stochastic scheduler which not only capable of accounting for schedule optimisation for conditionally simulated orebodies, in
Orebody Modelling and Strategic Mine Planning, second edition
the grade uncertainty in its formulation but also of managing the
(ed: R Dimitrakopoulos), pp 379-384 (The Australasian Institute of
risk associated with it in the deviation of production targets. Mining and Metallurgy: Melbourne).
The fact that a stochastic integer formulation solution of the Ramazan, S and Dimitrakopoulos, R, 2007a. Stochastic optimisation of
mine production schedule problem generates a higher value long-term production scheduling for open pit mines with a new
reflects the importance of incorporating geological uncertainty integer programming formulation, in Orebody Modelling and
into the scheduler problem formulation. Similar findings are Strategic Mine Planning, second edition (ed: R Dimitrakopoulos),
presented by Leite and Dimitrakopoulos (2007), which tested a pp 385-392 (The Australasian Institute of Mining and Mettalurgy:
different stochastic scheduler using the same copper deposit. Melbourne).
Like other available past studies, both studies make clear the Ramazan, S and Dimitrakopoulos, R, 2007b. Production scheduling with
advantages of using a stochastic approach in the mine production uncertain supply: A new solution to the open pit mining problem,
Cosmo Research Report, volume 2, pp 257-294.
schedule formulation.
Ravenscroft, P J, 1992. Risk analysis for mine scheduling by conditional
simulation, in Transactions of the Institutions of Mining and
ACKNOWLEDGEMENTS Metallurgy, Mining Technology, 101:A104-A108
Whittle, J, 1988. Beyond optimisation in open pit design, in Proceedings
The work in this paper was funded from NSERC CDR Grant Canadian Conference on Computer Applications in the Mineral
335696 and BHP Billiton, as well NSERC Discovery Grant Industries, pp 331-337 (Balkema: Rotterdam).
239019. Thanks are in order to Brian Baird, Peter Stone and Whittle, J, 1999. A decade of open pit mine planning and optimisation –
Gavin Yates, as well as BHP Billiton Diamonds and, in The craft of turning algorithms into packages, in Proceedings 28th
particular, Darren Dyck, for their support, collaboration and International Symposium on the Application of Computers and
technical comments. Operations Research in the Mineral Industry, pp 15-24 (Colorado
School of Mines: Golden).
REFERENCES
Dimitrakopoulos, R, Farrelly, C T and Gody, M, 2002. Moving forward
from traditional optimisation: Grade uncertainty and risk effects in
open-pit design, Transactions of the Institutions of Mining and
Metallurgy, Mining Technology, 111:A82-A88.
HOME
Regional Three-Dimensional Modelling of Iron Ore Exploration

Targets
V Osterholt1, O Herod2 and H Arvidson3
ABSTRACT meaningful and practical way. Estimates of such potential

The Australasian Code for Reporting of Mineral Resources and Ore
mineralisation can be developed in a variety of ways.
Reserves, December 2004 (the JORC Code) deals with reporting Historically, BHP Billiton has generated these as a deterministic
exploration target size and type, which are expressed as ranges in order to non-spatial grade and tonnage estimate. In the presence of
properly convey the high level of uncertainty that exists early in the uncertainty, deterministic estimates do not deliver the required
project evaluation process. To inform investors of the significant potential exploration target ranges. They also have limited use in strategic
mineralisation in BHP Billiton’s Pilbara tenements and to enable strategic planning, particularly in bulk commodities like iron ore, where
planning, a rigorous modelling approach has been devised for such early multi-pit blending options are a key strategic driver.
stage projects.
An innovative approach to modelling scenarios that capture the
uncertainty range has been developed. The approach integrates drill holes, MODELLING APPROACH
mapping, geophysical surveys and the geologist’s interpretation. The
modelling workflow consists of two parts: geology modelling, and grade
modelling. An implicit 3D modelling approach is utilised to rapidly A holistic view of geological uncertainty requires capturing the
generate multiple regional models of stratigraphy. Mineralisation envelopes uncertainty about the key factors controlling the grade-tonnage
and weathering horizons, providing further controls on the target range, are distribution of a prospective area. A target evaluation project
also modelled using fit-for-purpose tools and data structures. Grades are consists of three stages:
estimated with a recoverable resource estimation approach. The estimation
accounts for uncertainty of local grade distributions and provides 1. development of geology and grade scenarios;
multi-variate grade models that honour correlations between variables. The 2. geology modelling to build 3D representations of stationary
final product delivers a rigorous approach to the estimation of tonnage and
grade ranges for reporting the exploration target that captures the
domains, that is volume models; and
uncertainty of the estimate. 3. grade modelling to populate volumes with the relevant
grade estimates.
INTRODUCTION The initial conceptual analysis is based on understanding the
In the Hamersley Ranges, bedded iron ores comprise the bulk of states of knowledge (facts, assumptions, interpretations and
the area’s iron enrichment, varying between low phosphorous (P) unknowns), followed by describing the linkage between key
martite-microplaty haematite and medium-high P martite- influences on mineralisation. Concepts are then distilled into a
goethite or martite-ochreus goethite ores, originating from range of possible scenarios, a best-guess scenario (most likely)
enriched banded iron-formation (BIF) sequences of the and alternative pessimistic and optimistic scenarios (less likely
Brockman and Marra Mamba Formations. Stratigraphy is folded, but still reasonably likely as an outcome). These scenarios are
thrusted and faulted to various degrees across the Hamersley documented in a case matrix. This case matrix is the driving tool
Ranges as a result of multiple deformation phases. Structural of the modelling process – it links data analysis, interpretation
features often control the quantity and quality of mineralisation and scenarios developed during the conceptual analysis. All
by influencing the available volumes of BIF as well as the relevant details for all key drivers for the model cases are listed
enrichment processes. Figure 1 shows an example geological in a transparent and auditable way. A simple example illustrating
section through an openly folded orebody, with mineralisation the principle is given in Figure 2.
boundaries subparallel to stratigraphy and high continuity Developing scenarios and assigning probabilities is an expert-
throughout the syncline. Weathering effects often truncate driven process including team workshops to utilise all available
potentially economic mineralisation near the surface. Data sources of knowledge. Detailed data analysis and a review of
available during early evaluation work is usually sparse, with previous work and reports is combined with regional geological
only some drill holes per kilometre along strike, geological knowledge and experience from existing mining operations, which
surface mapping and geophysical data. Often data is historic and are then used to develop a range around the key influences.
associated with data entry errors, low precision, old drilling Examples include mineralisation in particular units of the
techniques and can be incomplete. Such data sets are not stratigraphic sequence, occurrence of (un-) favourable structures,
considered sufficient to support public reporting of mineral continuity of the mineralisation along strike and down-dip and
resources, which is an expression of the inherent uncertainty. In weathering intensity. The necessity and challenge of such expert
order to advance exploration and strategic planning, this judgements for early stage resource uncertainty assessment was
uncertainty around exploration targets needs to be addressed in a raised previously (eg Caumon and Caers, 2004). An innovative
approach to modelling these scenarios in three dimensions, with
the uncertainty, has been developed – it integrates drill holes,
1. MAusIMM, Senior Resource Geologist, BHP Billiton Iron Ore, mapping, geophysical surveys and the geologist’s interpretation.
PO Box 7122, Cloisters Square, Perth WA 6850. This paper describes the modelling approach in a generic way.
Email: [email protected] Results presented in figures provide an illustration of typical study
2. MAusIMM, Senior Consultant, SRK Consulting, PO Box 943, outcomes. The workflow consists of two parts, geology modelling
West Perth WA 6872. Email: [email protected] and grade modelling.
3. MAusIMM, Principal Resource Geologist, Resource Modelling These are presented in the next two sections, then the resulting
Group, BHP Billiton Iron Ore, PO Box 7122, Cloisters Square, modelling capabilities are summarised, and finally, the strengths
Perth WA 6850. Email: [email protected] of the approach and challenges are discussed.
V OSTERHOLT, O HEROD and H ARVIDSON
FIG 1 - Geological section of brockman iron-formation hosted orebody (from unpublished internal BHP Billiton report).
FIG 2 - Influence diagram with key parameters (yellow) (LHS) and case matrix of the potential mineralisation of a region (RHS).
MODELLING OF GEOLOGY mineralisation are triangulated to construct 3D domain models for

use in grade estimation. Using traditional mine planning software,
The basis for spatial modelling of potential mineralisation is the the time to construct and subsequently update these models can be
geology, which in most non-primary mineral systems can be significant and results in single geological models that are updated
expressed as the underlying lithology, with some form of annually for reporting purposes. When considering exploration
mineralisation system superimposed. In data-rich (typically well areas, the modelling time can often increase as additional non-data
drilled) areas, sectional interpretations of the lithology and constrained interpretations are required to produce appropriate
REGIONAL THREE-DIMENSIONAL MODELLING OF IRON ORE EXPLORATION TARGETS
triangulation models. To address the range of possible geological • reviewing consistency between different data types,
interpretations at exploration stages, a different approach to
geological modelling is needed. Implicit modelling tools are • introducing additional interpretation data (ie limited control
capable of rapidly building geologically robust triangulation sections), and
models of lithology and form the basis of a geological modelling • simplifying the interpretation in data rich areas.
workflow presented in Figure 3. Modelling zones of potential
mineralisation forms the second part of the workflow, including Through all the modelling, particular attention needs to be
combining data (where available) with a method of representing placed on the scale of the data with respect to the scale of the final
the understanding of the mineralisation processes. model. It is not possible to precisely fit the triangulations to very
closely spaced data or detailed interpretations in structurally
complex areas without post-modelling manual adjustment. In all
Modelling of lithology cases where these modelling approaches have been applied, the
Projects at exploration stages typically have surface geological conceptual understanding of the geology has improved, as the
mapping, comprising lithological contacts and orientation data and geologist can quickly explore the implications of different
some often shallow drilling. The project geologist will have a interpretations. By applying different geological interpretations
conceptual understanding of both the spatial distribution of the derived by an expert panel (eg increased fold amplitude or reduced
lithology and controls on the mineralisation system. The implicit throw on a fault), a series of lithological models can be produced
geological modelling tools used in this study combine a set of relatively quickly. These can then be used as the basis for
geological rules (in the form of a stratigraphic sequence) with mineralisation modelling and subsequent grade estimation.
geological interface and orientation data to construct a series of
3D functions which can be evaluated to create triangulations (see Modelling the mineralisation
Calcagno et al, 2008 for details of the methodology). The tools are
particularly appropriate for layered stratigraphy and thus very Within the lithology models, it is necessary to identify zones of
applicable for modelling banded iron ore deposits. (potential) mineralisation. In mapped areas, mineralisation may
The advantage of using implicit modelling tools is that they are have been identified at surface, and assay results from drilling
quick, with triangulations being generated within a few hours. will clearly indicate subsurface mineralisation. However, away
However, they do not represent a silver bullet to geological from this data, a method was required to highlight areas of
modelling. Input and control by a geologist needs to be applied to potential mineralisation based on an understanding of the
assure reasonable results. This typically takes the form of: mineralisation process. This is achieved through a combination
FIG 3 - Flow diagram showing a standard geological modelling workflow for resource estimation compared to a geological modelling
workflow using implicit modelling and separate mineralisation potential modelling to generate a series of triangulation models.
of suitable estimation of data in a 3D grid model and selecting 2. Conditional Simulation (CS) of geological and chemical
zones of the model based on modelled geological features. A components of an orebody on a fine grid of sample support
mineralisation indicator was applied to the drilling data values that are reblocked (averaged) to achieve SMU
(continuous values) to combine it with mapping (mineralised, support (Journel and Kyriakidis, 2004; Journel and
un-mineralised or unknown, ie binary or null). The parameters of Huijbregts, 1977).
the estimation of the combined data (distance, direction, number
of data required and method of estimation) was determined by an UC is based on the assumptions of the underlying Discrete
expert panel and varied depending on the scenario. By using Gaussian Model (DGM). The DGM is a method in which the
deformable 3D grids controlled by the lithology model, the original distribution is transformed to a standard normal
estimation was able to honour the structural complexity of the distribution using Hermite polynomials. This type of polynomial
stratigraphy. Away from the data, potentially mineralised areas of has the remarkable property of being ‘orthogonal’ for Gaussian
the grid model were selected by application of a conceptual values. This means that the transformation of the distribution for
mineralisation model. A series of 3D geographical information a variance correction can be simply made by applying a
system (GIS) tools were used which could query regions of the correction factor to the Hermite coefficients. Thus, there is no
model, eg those regions within 100 m of faults or occurring in assumption of normality or log-normality in the original data.
synformal fold hinges. The specific mineralisation rules were There is a reliance on the bi-normality of the transformed data
part of the expert panel scenario. The initial spatial visualisation and expected SMU values. In UC, the local panel’s mean and
of these ‘rules’ often identified unrealistic assumptions, which variance are estimated using kriging, while all other distribution
could then be re-evaluated by the geologists involved. moments are based on the global data distribution and honour
Cartier’s relationship. Hence the proportional effect is poorly
Combining the two approaches gives a grid model with a
captured by traditional UC. The method is fast and practical – a
series of mineralised zones – the zones can be triangulated for
multivariate implementation was recently provided (Deraisme,
visualisation purposes. The application of indicator estimation
Rivoirard and Carrasco, 2008). This approach enables a change
and a series of mineralisation ‘rules’ provides a repeatable and
of support for a primary variable and a set of secondary
quick method of generating zones of potential mineralisation
variables. The full pdf, F(X), of a primary variable, X, is
which are controlled by the underlying lithology model. For each
available for reporting at various cut-offs. Secondary variables,
lithology model, a series of mineralisation models could be built
Yi, can be reported as the expectation value of the conditional
representing a true range in possible geological scenarios.
pdf, E[F(Yi|X)].
Additional to the assumptions of the univariate DGM, the
GRADE MODELLING multivariate DGM assumes multi-Gaussian behaviour of the
variables considered. MIK is a method with little assumptions
Mineralisation grades add another level of uncertainty of potential regarding the shape of the grade distributions. The estimation of
mineralisation and have material impact on strategic scheduling the point-support pdf is local, with variance and higher moments
and blending studies. Statistics such as grade histograms and derived by kriging from the search neighbourhood. The
variograms are difficult to infer from sparse data and are therefore application requires extensive variography even for the univariate
uncertain. Grade uncertainty ranges are addressed using a case. Various support correction methods are available to infer
bootstrapping technique for resampling histograms. Variogram local SMU support pdfs. Neither MIK nor UC incorporate global
uncertainty was investigated, but was found to have only a minor or local uncertainty.
impact on the global grade tonnage distribution. Previous work on The common and commercially available CS techniques are
the topic includes Berckmans and Armstrong (1999). They report based on second order moments and also make assumptions
on a related but global approach for grade-tonnage estimation regarding the higher moments of the distributions. The
accounting for grade uncertainty. The approach presented here multi-Gaussian model is most widely used in the industry and is
aims at local grade modelling, so grade estimation will be fully defined by the Gaussian transformation of the sample
reviewed in the next few paragraphs before presenting the histogram and the related variogram model. An acceptable set of
approach in the following section. CS realisations should reproduce those two statistics within
Fit-for-purpose grade estimation has to consider the fact that ore ergodic fluctuations (eg Goovaerts, 1997). Some multivariate CS
is mined out as selective mining units (SMUs). The SMU algorithms suffer from some problems, for example variance
describes a defined volume that can be selected during mining and reproduction (Babak and Deutsch, 2008). CS models capture
is primarily linked to mining equipment. Typically, SMUs are grade variability and local uncertainty, enabling their quantitative
vastly larger than drill hole samples (eg Journel and Kyriakidis, evaluation within the CS model framework. These qualities cannot
2004), but in relation to exploration drill hole spacing, they are be delivered by estimation techniques. In general, CS has
small. When data is sparse, linear in situ resource estimation likely additional computation, storage and people requirements that raise
results in over-smoothed grade-tonnage curves. Such models the question of cost/benefit balance. New efficient algorithms
represent a too-low selectivity equivalent to excessively large retain SMU grades only and enable some CPU and storage savings
SMU size. A non-linear recoverable resource estimation approach (eg Boucher and Dimitrakopoulos, 2009; Deraisme, Rivoirard and
(or alternatively, a simulation approach) is required to enable Carrasco, 2008). However in this regional context model size
unbiased grade-tonnage curve estimation. The recoverable remains challenging even in terms of SMUs (eg for a 1 B tonnes
resource models that are generated form the basis of a realistic target – 5 M SMUs × 5 variables × 4 bytes × 50 realisations =
prediction of minable reserves that take dilution/ore loss, 5 GB).
economic factors, mining policies, etc into account.
Two approaches for recoverable resource estimation are LOCAL CHANGE OF SUPPORT
traditionally considered in the literature:
The proposed approach local change of support, or LoCoS,
1. Derivation of the sample support distribution (pdf) for a stands in between UC, MIK and CS and combines the strengths
panel and application of a change of support correction, ie of the three methods while avoiding some of their practical
conversion of sample support histogram to SMU support limitations in this large scale context. Given a panel within a
histogram. Methods include uniform conditioning (UC) and domain, three steps are performed:
multiple-indicator kriging (MIK) combined with support 1. kriging – to derive a local weighted histogram,
correction (eg Rivoirard, 1994; De-Vitry, Vann and
Arvidson, 2007). 2. spatial bootstrap of the local histogram, and
3. change of support for each histogram realisation from the negative kriging weights. Negative weights are tracked to enable
second step. validation of the estimation performance. The locations of the
samples in the local neighbourhood define data sparsity and are
Figure 4 illustrates the high level process flow for the LoCoS utilised to describe data redundancy in the spatial bootstrapping
grade modelling, showing decision points that occur when data process.
quantity is challenged, ie what to do if there is insufficient data
from which the distributions and spatial continuity can be
reasonably characterised. If data is too sparse, the local approach Spatial bootstrap
suggests more information than is available and becomes The randomisation is performed to account for the grade
meaningless. Global assessment of grades and grade uncertainty is histogram uncertainty given sparse data, ie the fact that sparse
more appropriate under such conditions. Given the regional data contains limited information regarding the true local pdf.
geological understanding and extensive history of iron ore This step is performed using a spatial bootstrap method based on
exploration and mining in the Pilbara, it is expected that analogous the Gaussian transformation of the data pdf and unconditional
deposits can be sourced to provide parameters (eg global pdf and simulations of a Gaussian random field at the sample locations
variogram models), even in the most data challenged situations. (Deutsch, 2007). While classical bootstrapping assumes that data
is independent from one another, the spatial bootstrap method
accounts for the auto-correlation of spatial data. As two
Select panel
auto-correlated samples contain information regarding each
other’s value, there is effectively less independent data when the
data is auto-correlated.
The spatial bootstrap method was adjusted to the requirements
Sufficient No of assessing local multivariate pdf uncertainty:
local data?
• The bootstrap is based on the primary variable and captures
data redundancy due to spatial correlation of this variable.
Sufficient No
Yes global • The original method is univariate. For this approach, the
Kriging for local pdf data? complete sample records are carried through to the resampled
Data analogue data set to generate multivariate realisations. The complete
Yes sample records maintain the experimental multivariate
Spatial bootstrap Build global pdf relationships of the data.
• The realisations of the pdf are ranked based on data statistics
such as the mean grade of the primary variable – grade
Change-of-Support models capturing bias scenarios can be generated from
equally ranking local pdfs.
Build parcel model The accumulation of uncertainty from multiple equally
ranking local pdfs was found to quickly exceed uncertainty
FIG 4 - Process flow diagram for the grade modelling approach. ranges based on the global data. To keep grade ranges reasonable
and aligned with the ranges specified in the case matrix the
number of realisations per panel can be kept relatively small.
Kriging Typically about ten realisations for each panel accumulate to an
uncertainty range comparable to 1000 global realisations (Figure 6).
The kriging provides weights for the samples in the search
neighbourhood. Ordinary Kriging (OK) weights, accompanied Change of support
by primary and secondary assay variables, define an
experimental multivariate sample-support distribution for the After the (optional) kriging and spatial bootstrap, a weighted
panel (Figure 5). Negative OK weights are set to zero and the point-support multivariate pdf is available for the panel, with
remaining weights restandardised to sum to 1. The search multiple pdf realisations capturing local grade uncertainty. These
criteria, such as minimum required number of samples, have to point-support panel pdf realisations require change of support
balance the resolution of the final pdf and the potential for correction to derive the SMU-support grade tonnage distribution
OK for panel pdf for panel
100
ë4
ë5
ë1 ë5 ë1
P
ë2 ë4
ë2
ë3
ë3
s3 s2 s1 s5 s4
FIG 5 - Ordinary kriging for panel gives weights λi for samples si in neighbourhood to derive local weighted histogram.
display the model averages for each bin. The variance of both
variables is reduced in the SMU support with overall
reproduction of the conditional mean of Fe. Some grade parcels
are generated outside of the point scatter due to minor local
difficulties with the change of support parameters. Figure 10
summarises the grade-tonnage curves from nine model scenarios
showing three clusters in the tonnage and the grade plots
identifying the different geology scenarios. Grade uncertainty is
considerable, but mostly driven by geology and the associated
declustering effect. Grade uncertainty is a major contributor to
tonnage uncertainty at high Fe cut-offs.
CONCLUSIONS
The modelling process allows rapid generation of alternative 3D
models of exploration scenarios. The process is data-based and
transparently includes subjective expert judgements where
required. Implicit geological modelling tools enable the
integration of a range of geological data types and the quick
FIG 6 - Cumulative histogram of Min, Mid and Max bootstrap development of multiple lithological models. By basing the
scenarios versus data for Al2O3 – based on 500 realisations estimation of potential mineralisation indicators on these models,
(global bootstrap) and ten realisations per panel (local bootstrap). the resulting volumes are consistent with drill hole data, mapping
and the structural framework. The grade modelling approach
presented delivers multivariate models of recoverable resources
for the panel. The change of support is based on the multivariate that reflect data statistics, including spatial trends. These models
discrete Gaussian model (reviewed above – Deraisme, Rivoirard are a realistic representation of deposit to regional scale grade
and Carrasco, 2008). In the LoCoS algorithm, a Gaussian tonnage curves and avoid excessive smoothing, the typical
anamorphosis is performed on each of the pdf realisations that shortcoming of linear interpolators where data is sparse. Uni- and
were obtained by the spatial bootstrap process, hence reducing bi-variate data statistics are well captured with the proposed
the strict assumption about the global distribution required in LoCoS method.
traditional UC. This results in better representation of Due to the bi-variate nature of the change of support approach,
proportional effect by LoCoS. Then the change of support is applying cut-offs is limited to a single variable. Using any
performed with variance correction factors derived from the secondary cut-off variable would mean changing the selectivity for
LMC. that variable, as parcels with variable tonnages had to be selected,
with volumes generally exceeding the SMU size. In practice, this
RESULTS demands an upfront decision regarding the main grade control
variable. An alternative approach could be formulated as a
The process generates a set of 3D models reflecting the range of variation of uniform conditioning (UC), with global uncertainty
grade-tonnage uncertainty for the model area. Figure 7 shows being captured by utilising multiple Gaussian anamorphoses. This
three scenarios generated for an area of approximately 25 cubic approach would be easy to implement with UC with only a single
kilometres. The continuity of mineralisation (coloured on brown step additionally performed, the spatial bootstrapping of the global
topography) increases from the pessimistic to the optimistic data. However, the assumption of a known (estimated) local panel
scenarios. The potty volumes around some drill holes are not grade would result in understating grade and tonnage uncertainty.
ideal, but the resulting global tonnage ranges are fit-for-purpose. The 3D geology models add value by enabling visualisation of the
The kriging of panel grades can be evaluated using the same scenarios, thus developing further learning. They also have current
statistical and graphical tools as any kriged estimate. Figure 8 application in planning of exploration programs. The exploration
shows the Eastings swath plots for Fe of blocked sample data target model scenarios allow global grade tonnage reporting and
(red) and estimated blocks (blue). Overall grade fluctuations are strategic planning. Future anticipated uses of either the current
captured in the model with some spatial smoothing. geology models, or the geological modelling workflow, include
Figure 9 shows the resulting scatter plot of Fe versus Al2O3 for the wider mining related geosciences, such as geotechnical
SMU support (black) versus samples (pink). The red points engineering and hydrogeology.
Pessimistic Best Guess Optimistic
FIG 7 - Example of three mineralisation scenarios in plan view.
ACKNOWLEDGEMENTS
The authors wish to thank an anonymous peer reviewer for
constructive review of this paper, Sia Khosrowshahi (Golder
Associates) for his help in developing ideas for the grade
modelling approach and Richard Mattsson (BHP Billiton) for
practical guidance and motivation. Also thanks to BHP Billiton
Iron Ore for permission to publish this work.
REFERENCES
Babak, O and Deutsch, C V, 2009. Collocated cokriging based on
merged secondary attributes, Mathematical Geosciences, volume
41(8)921-926.
Berckmans, A and Armstrong, M, 1999. Bootstrapping: A fast way to
simulate QTz curves, Mathematical Geology, 31(4):471-485.
Boucher, A and Dimitrakopoulos, R, 2009. Block simulation of multiple
correlated variables. Mathematical Geosciences, 41(2), pp 215-237.
FIG 8 - Eastings swath plot validation for mean panel Fe grades. Calcagno, P, Chiles, J P, Courrioux, G and Guillen, A, 2008. Geological
modelling from field data and geological knowledge: Part I,
Modelling method coupling 3D potential-field interpolation and
geological rules, Physics of the Earth and Planetary Interiors,
171(1-4):147-157.
Caumon, G and Caers, J K, 2004. Assessment of global uncertainty for
early appraisal of hydrocarbon fields, Society of Petroleum
Engineers, paper 89943.
Deraisme, J, Rivoirard, J and Carrasco, P, 2008. Multivariate uniform
conditioning and block simulations with discrete gaussian model:
Application to Chuquicamata deposit, in Proceedings Eighth
International Geostatistics Congress (eds: J M Ortiz and X Emery),
pp 69-78.
Deutsch, C V, 2007. A statistical resampling program for correlated data:
Spatial_boostrap, paper 401, in Annual Report Centre for
Computational Geostatistics, Report 6.
De-Vitry, C, Vann, J and Arvidson, H, 2007. A guide to selecting the
optimal method of resource estimation for multivariate iron ore
deposits, in Proceedings Iron Ore 2007, pp 67-78 (The Australasian
Institute of Mining and Metallurgy: Melbourne).
Goovaerts, P, 1997. Geostatistics for Natural Resources Evaluation,
Applied Geostatistics Series, 483 p (Oxford University Press:
FIG 9 - Bivariate validation of Fe versus Al2O3; pink – data, black – New York).
SMU, red – average SMU. Journel, A G and Huijbregts, C J, 1977. Mining Geostatistics, 600 p
(Blackburn Press: Caldwell).
Journel, A G and Kyriakidis, P C, 2004. Evaluation of Mineral Reserves:
A Simulation Approach, 216 p (Oxford University Press: New York).
Levy, B and Mallet, J L, 1999. Discrete smooth interpolation: Constrained
discrete fairing for arbitrary meshes [online], GoCAD whitepaper.
Available from: <http://alice.loria.fr/publications/papers/1999/smooth
ing/smoothing.pdf>
Mallet, J L, 1992. Discrete smooth interpolation, Computer Aided Design
Journal, 24(4):263-270.
Nowak, M and Verly G, 2007. A practical process for simulation, with
emphasis on gaussian simulation, in Orebody Modelling and
Strategic Mine Planning, second edition (ed: R Dimitrakopoulos),
pp 69-78 (The Australasian Institute of Mining and Metallurgy:
Melbourne).
Rivoirard, J, 1994. Introduction to Disjunctive Kriging and Non-Linear
Geostatistics, 180 p (Clarendon Press: Oxford).
FIG 10 - Grade tonnage curves for nine model scenarios.
HOME
Conditional Simulation of Grades Controlled by Geological

Indicators
C Xu1 and P A Dowd2
ABSTRACT important controlling factor in ore reserve estimation and

selective mining.
This paper describes a method of conditionally simulating an orebody in
which geological indicators have a controlling influence on the Exploration and development drilling was originally on a
occurrence and spatial distribution of the grades of primary economic 30 m × 30 m grid with some areas drilled at 30 m × 15 m and
variable. The simulation comprises two components: 1. grades, and 2. peripheral areas drilled at 60 m × 60 m, one axis of the grid
geological controls. having a parallel orientation to the major strike direction.
In the example provided in the paper, the authors use plurigaussian Geostatistical studies indicated an optimal drilling grid of
simulation to simulate the occurrence of quartz veins and sequential 25 m × 12 m, but a cost-imposed compromise increased the
Gaussian simulation to simulate the associated gold grades. The minimum dimensions to 30 m × 15 m. The holes were drilled
simulated orebody can then be sampled at various scales to yield from the surface at a dip of approximately 40° to the north or
estimates of gold content and associated geological indicators. Estimates south to intersect the predominantly vertical quartz veins at as
derived from the data at each sampling scale can be compared against the
great an angle as possible, as illustrated schematically in Figure 1.
complete simulated orebody to assess the effects of estimated indicators
on recoverability. The cost of additional sampling/drilling can be assessed Initial drilling was by direct circulation and thereafter by reverse
against the increased profit generated by more accurate prediction of the circulation. Hole diameters varied from 90 mm to 150 mm and
geological indicators. were eventually standardised at 135 mm.
INTRODUCTION Relationship between quartz and gold

Geological indicators are an important element in the selective The correlation coefficient between the quartz and gold contents
open pit mining of many types of orebodies. A good example is of samples is not significant even though there is strong
provided by quartz-vein controlled gold orebodies in which geological and mineralogical evidence for the association of gold
selective mining units are defined by outlines around visible with quartz veins. This lack of statistical correlation may be
occurrences of quartz veins and/or natural fractures on exposed caused by the confinement of the association to the contact zones
benches. Selection can be implemented in this way only when a between quartz veins and granodiorite. If this is the case, the
mining bench is exposed and even then only exposed veins and important factor would be the number of such contacts or veins
fractures can be used in selection. It is much more difficult to within a given sample rather than the total amount of quartz in a
include the effects of such selection processes in recoverable sample, as illustrated in Figure 2. This information is not
reserve estimates. The ability to predict accurately the available from percussion holes. Another explanation may be
occurrence of geological indicators and their effects on ore that, although gold is associated with quartz, the mineralisation
boundaries largely determines the amount of ore loss and ore may have independent controls within the quartz-vein system.
dilution; minimising each of these is a major determinant of Note, however, that there is no difference in the spatial variability
profitability. of gold directly associated with quartz and that of gold not
Problems associated with identifying geological indicators and directly associated with quartz as shown by the two downhole
their relationship to ore zones are exacerbated in low-grade gold variograms in Figure 3. These variograms can be modelled as
orebodies. This is especially so when the only samples available spherical models with a nugget variance and two structures – the
are from percussion drilling, from which it is not possible to first with a range of 4 m and the second with a range of 20 m.
discern the detailed nature of the geological indicator. The complete three-dimensional model is given in Table 3.
Figure 4 presents log-probability plots of the cumulative
THE OREBODY histograms of gold assays of samples within a representative
zone of the orebody. The Figure shows the cumulative histogram
A full description of the geology and mining operation can be of the gold assays of all samples with corresponding quartz
found in Dowd (1995). Gold mineralisation occurs within a assays of zero wt per cent together with the cumulative
network of steeply dipping quartz veins in the contact between histogram of gold assays of samples with corresponding quartz
older granodiorite and limestone/acid volcanic rocks. The gold assays greater than zero wt per cent. In spite of the geological
has an erratic distribution and is mainly concentrated in and evidence, the only significant difference between the two
around high-grade quartz veins. Granitic zones between the
log-probability plots is the mean grade of the two sets of values.
high-grade veins contain fine-grained gold associated with
pyrite. The difference between the mean of the gold values associated
with quartz and of those not associated with quartz is largely due
The geological relationship between quartz and gold is a
to a few high-grade samples. A summary of statistics for the
fundamental element in the selective mining of ore and the final
delineation of ore zones is achieved largely by mapping quartz representative zone is given in Table 1.
veins on exposed benches. This relationship is, thus, an The histograms in Figure 4 and the statistics in Table 1
illustrate a typical problem in the evaluation of gold orebodies –
the presence of a small number of high-grade outliers. The
1. Senior Lecturer, School of Civil, Environmental and Mining maximum grade in the data set that was used to calculate the
Engineering, Engineering North Building, The University of statistics in Table 1 was 4360 g/t. If the high-grade value of
Adelaide, SA 5005. Email: [email protected] 4360 g/t is excluded from the calculation, the mean of the
2. FAusIMM, Executive Dean, Faculty of Engineering, Computer and quartz-associated gold assays is 3.45 g/t and the variance is
Mathematical Sciences, The University of Adelaide, SA 5005. 824.4 (g/t)2. The mean and variance can be further reduced by
Email: [email protected] excluding two or three other high-grade outliers.
C XU and P A DOWD
FIG 1 - Schematic of typical cross-section.
assay. As an example, consider the entry for gold assays greater

quartz than or equal to 0.5 g/t. There are 644 samples with zero wt per
(a) (b)
cent quartz and a gold assay greater than or equal to 0.5 g/t.
quartz
Of these, 39 per cent are within an average of 1 m of a non-zero
quartz assay, 58 per cent are within 2 m of a non-zero quartz
FIG 2 - Possible relationships between gold content and numbers assay and 68 per cent are within 3 m of a non-zero quartz assay.
of quartz-granodiorite contacts. Examples of drill hole These distances are measured along the drill hole, so that if
intersections: (A) with low quartz content and high number of quartz is encountered within 1 m of a sample, it occurs in the
quartz-granodiorite contacts and (B) with high quartz content and contiguous sample on either or both sides. Distances of 1 m, 2 m
low number of quartz-granodiorite contacts.
and 3 m along a drill hole are equivalent to horizontal distances
of 0.7 m, 1.4 m and 2.1 m, respectively.
The following observations can be made on the basis of the
information summarised in Table 2:
• a nearly constant 40 per cent of all significant gold assays
that are not associated with quartz occur within 1 m of a
quartz occurrence, ie 60 per cent are further than 1 m (0.7 m
horizontally) from a quartz vein;
• a nearly constant 60 per cent of all significant gold assays
that are not associated with quartz occur within 2 m of a
quartz occurrence, ie 40 per cent are further than 2 m (1.4 m
horizontally) from a quartz vein; and
• a nearly constant 70 per cent of all significant gold assays that
are not associated with quartz occur within 3 m of a quartz
occurrence, ie 30 per cent are further than 3 m (2.1 m
horizontally) from a quartz vein.
The mean gold grade of samples that have a zero quartz assay
and a significant gold assay and which are within 1 m of a quartz
vein (column 5 in Table 2) is consistently and significantly less
than the mean gold grade of all samples with zero quartz assay
(column 3).
FIG 3 - Downhole variograms of the logarithms of gold grades of The three observations appear to be evidence that significant
samples in which no quartz is recorded and for samples in which gold mineralisation is not directly associated with quartz veining
quartz is recorded. and/or that there is significant sample contamination, which is
often a problem in percussion drilling, especially with direct
circulation. If contamination is a factor, the nearest occurrence of
It is possible to fit three lines of slightly different slopes at quartz for any sample should, more often than not, be further up
values of 1.0, 0.01 and 0.001 g/t on the cumulative histograms in the hole. As there is no evidence of this, the only possible
Figure 4. The 1 g/t threshold appears to be related to gold- conclusion is that, although the coarsest-grained particles of gold
bearing quartz-vein occurrences, although the relationship cannot may be associated with quartz contacts, there is a significant
be quantified. proportion of finer-grained gold that is not directly associated
with quartz veins.
A further study was conducted to determine whether samples
with a significant gold assay and a zero quartz assay occur close Extending the calculations shows that approximately 80 per
to the occurrence of quartz. A summary of the study is given in cent of all samples with zero wt per cent quartz and significant
Table 2, which was constructed by selecting all samples with a gold grades are within 5 m of a quartz occurrence. A downhole
quartz assay of zero wt per cent and then searching for the distance of 5 m corresponds to a horizontal distance of
nearest sample in the same drill hole with a non-zero quartz approximately 3.5 m. When all samples are considered (ie
CONDITIONAL SIMULATION OF GRADES CONTROLLED BY GEOLOGICAL INDICATORS
100
10
Gold grade, g/t, of 1 m drill hole samples
Quartz > 0% Quartz = 0%

1.0
0.1
0.01
0.001
0.01 0.05 0. 1 0.2 0.5 1 2 5 10 20 30 40 50 60 70 80 90 95 98 99 99.8 99.9 99.99
Cumulative frequency
FIG 4 - Log-probability plots of gold grades for samples containing no quartz and samples containing quartz.
TABLE 1 shown in Figure 5 in which a spherical model with a range of

3 m has been fitted. This range is very similar to the range of the
Statistics of gold assays with and without associated quartz
(subvolume).
first structure observed on the downhole variograms for gold
(Figure 3 and Table 3) and may reflect the controlling influence
Statistic Quartz = 0 wt% Quartz >0 wt% of quartz veins on gold occurrence.
(n = 3849) (n = 1856) The challenge in the simulation of the gold grades is to
Mean, g/t 0.95 5.79 reproduce the observed associations with quartz in the absence of
any significant statistical correlation and any spatial covariation
Variance, (g/t)² 55.3 11 050.1
in the form of cross-variograms. Independent univariate
Median, g/t 0.06 0.30 simulation of gold and quartz grades will reproduce the spatial
25th percentile, g/t 0.02 0.08 correlation of each variable but not the physical association
described above. Standard forms of co-simulation of gold and
75th percentile, g/t 0.25 1.30
quartz do not help either because of the negligible spatial
90th percentile, g/t 1.00 5.00 cross-correlation between the two variables. We propose a
modified version of sequential Gaussian simulation which uses
an adaptive local mean during the simulation. The adaptive local
including samples that contain quartz) 72 per cent of all gold
mean in this case is controlled by previously simulated quartz
assays greater than or equal to 0.5 g/t are either directly associated grades.
with quartz or are within 1 m of a quartz occurrence; 80 per cent
are either directly associated with quartz or are within 2 m of a As a preliminary test for the proposed method, we use only the
eastern part of the orebody. In this subset, there is a total
quartz occurrence; and 85 per cent are either directly associated
of 16 626 samples of which 31 per cent (5077) record quartz
with quartz or are within 3 m of a quartz occurrence. These occurrence. The association between gold and quartz can be
percentages increase as the cut-off grade increases because a demonstrated by plotting the average gold grade of samples
larger proportion of high-grade gold is associated with quartz. against the distance of the samples to the closest quartz sample,
These observations have important implications for selective as shown in Figure 6a. Note that the average grade for zero
mining, in particular for optimum mining or selection widths. distance is the average gold grade for those samples that contain
The relationship of gold and quartz within distances of up to quartz. Average grades for other distances are the average gold
3 m is reflected in the downhole variogram for quartz grades grades for samples in which quartz is not present and they are
C XU and P A DOWD
TABLE 2
Proportion of samples with 0 wt per cent quartz within specified distance of quartz veins.
Gold, > g/t No of samples Mean of One metre Mean of Two metre Mean of Three metre Mean of
samples, g/t distance per samples, g/t distance per samples, g/t distance per samples, g/t
cent of total cent of total cent of total
0 3849 0.95 31 0.98 49 1.1 62 1.14
0.2 1079 3.24 39 2.7 58 3.24 69 3.47
0.5 644 5.23 39 4.27 58 5.16 68 5.61
1 379 8.39 37 7.1 55 8.6 68 9.06
1.5 299 10.31 38 8.57 56 10.36 70 10.82
2 244 12.25 40 9.69 57 12.09 72 12.58
2.5 207 14.03 40 10.98 56 14.12 72 14.45
3 169 16.58 41 12.62 59 16.09 73 16.82
3.5 142 19.11 40 14.58 58 18.75 73 19.46
4 134 20.02 41 14.97 58 19.52 74 20.25
4.5 121 21.72 41 16.05 60 20.8 74 21.85
5 114 22.75 42 16.53 60 21.74 74 22.86
5.5 102 24.82 44 17.28 63 22.77 74 24.94
6 96 26.01 44 18.1 63 23.61 73 26.59
7 82 29.35 44 20.04 63 26.58 72 30.35
10 55 39.71 38 28.34 58 37.98 67 43.49
The average wt per cent quartz corresponding to Figure 6a is

TABLE 3 given in Figure 6b. The figure further illustrates the lack of
Spherical model variogram used in the sequential Gaussian correlation between the gold and quartz grades.
simulation.
In this application, a univariate gold simulation will reproduce
Nugget Structure 1 Structure 2 the spatial correlation, but not the association relationship, as
shown in Figure 7.
C 0.41 0.35 0.24
a (range) - strike 30 30
COMBINED PLURIGAUSSIAN AND SEQUENTIAL
- down-dip 25 25 GAUSSIAN SIMULATION
- cross-dip 4 20
As quartz vein location is a controlling factor in gold occurrence it
is sensible to include the quartz vein information when simulating
the gold grade. The approach proposed in this paper involves two
distinct steps. The quartz vein is simulated first using the
plurigaussian method, which provides the basic structure for the
next step in the simulation. In the second step the gold grades are
simulated using the sequential Gaussian method but using adaptive
local means rather than the global mean. The local mean for each
Gaussian simulated value is obtained as a function of the shortest
distance from the simulation point to a previously simulated quartz
vein in a manner similar to that illustrated in Figure 6a. On this
basis the sequential Gaussian simulation reproduces the spatial
correlation of gold grades and the adaptive local mean obtained
from the simulated quartz model reproduces the gold-quartz
association relationship shown in Figure 6a.
Plurigaussian simulation (Le Loc’h and Galli, 1996) is a
versatile method for simulating geological structures. It truncates
multiple correlated Gaussian fields to produce indicator
variables, which represent the geological units to be simulated.
For details of the method used here and the open-source codes
for the implementation, readers are referred to Dowd (2003) and
Xu et al (2006).
FIG 5 - Downhole experimental variogram and model for As the actual quartz grade is not an important indicator for
quartz grades. gold concentration, only presence/absence of quartz is of interest
in simulating the quartz vein model for the final simulation. The
indicator for the presence/absence of quartz can be treated as a
calculated on the basis of the distances of the samples to the category variable and the plurigaussian method can then be used
closest occurrence of quartz. The relationship shown in Figure 6a to simulate it. Plurigaussian simulation in this case is a truncated
applies to the subset used in this study and could be replaced by univariate Gaussian simulation, as only one Gaussian variable is
a relationship relevant to an application. required to simulate one indicator variable.
Average Au (g/t) Average Qz (%)

20
1.6
16
1.2
12
0.8
8
0.4 4
0 0
0 10 20 30 0 10 20 30
(A) Distance to Qz vein (m) (B) Distance of Au sample to Qz vein (m)
FIG 6 - Average gold and quartz grades versus the gold sample distance to nearest quartz occurrence.
Average Au (g/t) a standard Gaussian variable corresponding to a cumulative

frequency of 0.309). Two variables are created to define the
1.6 spatial correlation of quartz veins for the plurigaussian
Traditional sgsim simulation. F1 represents the presence of quartz (ie F1 = 1 if
1.2 quartz >0 per cent and 0 otherwise) and F2 represents the
absence of quartz (ie F2 = 1 if quartz = 0 per cent, 0 otherwise).
0.8 The variogram of F1 and the cross-variogram F1-F2 are shown in
Figure 8. These variograms are used by PGSim1 (Xu et al, 2006)
0.4 to derive the spatial correlation model for the Gaussian field. The
optimal model found is an exponential model with a = 17 m.
0 This Gaussian model will produce the theoretical spatial
0 20 40 60 correlations for F1 and F2 shown as the solid lines in Figure 9.
Distance to the closest Qz sample (m)
Distance to the closest Qz sample (m)
Sequential Gaussian model
FIG 7 - Result from traditional SGSIM (Deutsch and
Journel, 1992). The sequential Gaussian method is used for the simulation of
gold grades. The variogram model for the gold grades is reported
Monte Carlo sampling is used to derive the adaptive local mean in Dowd (1995). The parameters of the model fitted to the
corresponding variogram of the normal score transform of the
to be used in the simulation of gold grades. The distribution used
gold grades are given in Table 3. This variogram model for the
for the Monte Carlo simulation is derived directly from the data strike-plunge direction is shown in Figure 10 together with the
set. For example, Figure 8 shows the histogram, derived from the experimental variogram calculated from the data and the
dataset, of the gold grades of samples that are 2 m away from the variogram calculated from a simulation.
nearest sample containing quartz. This histogram is used to
generate the adaptive local mean for the gold grade simulation at
those simulation locations that are 2 m away from the nearest SIMULATION RESULTS
simulated quartz occurrence. Figure 11 shows a simulated quartz vein model generated by the
plurigaussian simulation. This model is used, as described above,
No. Samples( 1000) to derive the adaptive local mean for the sequential Gaussian
4 simulation of the gold grades. A cross-section from one such
simulation is shown in Figure 12.
3 The spatial correlation models are reproduced satisfactorily by
the combined plurigaussian and sequential Gaussian simulation,
2 as is evident from Figures 9 and 10. Most encouragingly, by
using the approach described here, the reproduction of the
association between quartz and gold has been improved, as
1
shown in Figure 13. The improvement over models generated by
sequential Gaussian simulation without the control imposed by
0 the prior quartz simulation (Figure 7) is significant.
0.01 0.51 1.01 1.50 2.00 2.50
Histogram of A× grade for distance = 2 m CONCLUSIONS
FIG 8 - Gold grade histogram derived from data. This paper describes a simple two-step simulation method to
incorporate the controlling effects of quartz veins when there is
no significant direct correlation between gold and quartz. The
SPATIAL CORRELATION MODELS geological structure in this approach is simulated by the
plurigaussian method and the primary variable (gold) is
Plurigaussian model simulated by the sequential Gaussian method with an adaptive
local mean derived from the simulated geological structure. The
The proportion of samples in the dataset that contain quartz (ie improvement in terms of the reproduction of the association is
samples with quartz grade > zero per cent) is 0.309. This number significant. If diamond drill core data are available and logged
is used in the plurigaussian algorithm to determine the truncation it would be possible to simulate the detailed shapes of the veins
threshold t = -0.498 for the Gaussian variable (ie t is the value of rather than the simplistic simulation of quartz occurrence.
C XU and P A DOWD
γF1 γF1-F2 Distance (m)

0.25 0
0 100 200 300
0.2 -0.05
0.15 -0.1
0.1 -0.15
0.05 -0.2
Distance (m)
0 -0.25
0 100 200 300
Data Model Simulation
FIG 9 - Spatial correlation models for F1 and F2.
γ - normal scores
1.2
Data
1
Model
0.8
Simulation
0.6
0.4
0.2
Distance (m)
0
0 20 40 60 80 100
FIG 10 - Variogram of normal scores in the strike-plunge direction.
Quartz vein
North-south cross section @ easting=1808
3D quartz model within the defined orebody
FIG 11 - A simulated quartz model generated by plurigaussian simulation.
North-south cross-section @ easting = 1808

0.0 >0.5
FIG 12 - A simulation of gold grades generated by sequential Gaussian simulation.
Average Au (g/t) The work described in this paper is only preliminary and is
presented primarily as a test of the proposed method. The
1.6 encouraging results reported warrant further detailed
investigation for the entire orebody. More detailed modelling of
1.2 the plurigaussian model is also necessary.
One simulation
0.8 REFERENCES
0.4 Deutsch, C and Journel, A G, 1992. Geostatistical Software Library and
User’s Guide, 340 p (Oxford University Press).
0 Dowd, P A, 1995. Björkdal gold-mining project, northern Sweden,
Transactions of the Institutions of Mining and Metallurgy Mining
0 10 20 30
Industry, Mining Technology, 104:A149-A163.
Distance to the closest simulated Qz vein Dowd, P A, Pardo-Igúzquiza, E and Xu, C, 2003. PLURIGAU: A
computer program for simulating spatial facies using the truncated
plurigaussian method, Computers and Geosciences, 29(2):123-141.
FIG 13 - Results from the two-step simulation.
Le Loc’h, G and Galli, A, 1996. Truncated plurigaussian method:
Theoretical and practical points of view in Proceedings Fifth
One significant application of the simulations generated by International Geostatistics Congress (eds: E Y Baafi and N A
these methods is the optimal design of drilling grids to detect Schoffield), 1:211-222 (Kluwer Academic Publishers: Dordrecht).
geological indicators and to incorporate the effects of these Xu, C, Dowd, P A, Mardia, K V and Fowell, R J, 2006. A flexible true
indicators in recoverable reserve estimates. plurigaussian code for spatial facies simulations, Computers and
Geosciences, 32(10):1629-1645.
HOME
Mining Geostatistics with the Open Source SGeMS Software and

the GsTL Library
A Boucher1
ABSTRACT from its non-obtrusive graphical user interface and from the
possibility to directly visualise data sets and results in a full 3D
Stochastic representations of mineral deposits are complex and off-the
-shelf algorithms are not always readily suitable for their modelling. The
interactive environment. The second objective was to design
free and open source software SGeMS (Stanford Geostatistical Modelling software that would cater to the needs of power-users. In SGeMS,
Software) and the library GsTL (Geostatistical Template Library) with its most of the operations performed using the graphical interface can
collection of advanced probabilistic algorithms, provides convenient also be executed through scripts. The integrated support for the
resources for deposit and uncertainty modelling, research and algorithm Python scripting language enables the creation of macros for
development. SGeMS also encompasses the open source Python scripting simple tasks all the way to complete nested applications with an
language that allows further processing of geostatistical simulations such as independent graphical interface. New features can be conveniently
grade control. Finally, the latest multiple-point simulation algorithms are added into SGeMS through a system of plug-ins, ie pieces of
also available in the software package. software which can not be run by themselves, but complement a
This paper explores the capabilities of the libraries for mining main software. In SGeMS, plug-ins can be used to add new
geostatistics. Two examples are used to illustrate the potential of the geostatistics tools or utilities such as defining new file filters to
SGeMS software in developing practical tools. First, a C++ plug-in that import/export data into appropriate file format.
seamlessly integrates the Group Sequential Gaussian Simulation (g-sgsim)
into the software SGeMS is presented. Then a Python script is written to
select optimal variogram parameters through a cross-validation procedure. THE CASE FOR OPEN SOURCE SOFTWARE
SGeMS is free and open source. Open source software gives the
INTRODUCTION users access to the source code, which can be modified with or
Numerical 3D representations of mineral deposits with geo- without some limited restrictions, depending on the license
statistical techniques are widely used both in academia and in requirement. For example, SGeMS is distributed under a GNU
industry. The use of geostatistical modelling techniques in mining General Public License (GPL) and must always be distributed with
applications is only possible through efficient and adaptable its source code. Any changes or additions to the software are
software. Existing software specially designed for the mining automatically under the GPL license and the source code must be
industry has made the spread of geostatistical techniques in provided whenever it is distributed. See http://www.gnu.org/copy
modelling and engineering possible. However, mineral deposits left/gpl.html for more details on the GPL license. Open source
and other natural phenomena can be so complex that algorithms software gives flexibility and freedom to the users. The source
offered in off-the-shelf software may not be suitable without code can be modified to address complex problems that may not
significant modifications. There are two solutions: have been easily handled from the traditional interface of
geostatistical algorithms. Additionally, open source software can
1. either a software package offers an exhaustive series of grow through contributions from the user’s community. Access
algorithms, or to the source code also allows the users to track and correct bugs
without waiting for an official release or patches from a vendor.
2. it allows the user to have access to the code and add/modify
More importantly, the transparency permits the user to validate
algorithms.
the algorithms by inspecting the actual codes instead of using
For instance, the availability of the source code is a significant them as black boxes.
contributing factor for the popularity of the GSLIB suite of
algorithms (Deutsch and Journel, 1998).
This paper shows the capabilities of SGeMS (Remy, Boucher ADDING A GEOSTATISTICAL ALGORITHM TO
and Wu, 2009), an open source geostatistics software, for mining SGeMS
applications. SGeMS, fully written in C++, includes geostatistical
algorithms developed at the Stanford Center for Reservoir New algorithms can be seamlessly integrated in SGeMS through a
Forecasting (SCRF) at Stanford University. As of October 2008, it plug-in mechanism. Once loaded, all user-generated algorithms
has been downloaded more that 12 000 times (not necessarily share the same interface with SGeMS as the existing ones, and
unique users). It was written by Remy (Remy et al, 2002; Remy, have access to the same facilities that SGeMS provides, such as
2004a, 2004b), and is maintained by a group of developers at visualisation and data analysis. To facilitate the writing of
Stanford University and elsewhere. SGeMS was written with two algorithms, the user has access to all the libraries of SGeMS (such
goals in mind. The first goal, geared toward the end-user, was to as the search neighbourhood and sequential simulation) and of the
provide a user-friendly software which offers a wise selection of Geostatistical Template Library (GsTL), which contains geo-
geostatistics tools – ranging from commonly used variogram statistical functions and utilities such as probability distribution
based algorithm (eg kriging, cokriging, indicator kriging, functions, builders of kriging systems and variograms/covariances
sequential Gaussian simulation and cosimulation, indicator functions. Thanks to the generic implementation of these libraries,
sequential simulation and direct sequential simulation), to the the user can focus on the specificity of the new algorithm without
more recent training image-based (multiple-point statistics) having to rewrite basic functions that are common to most geo-
simulation (Strebelle, 2002; Wu, Boucher and Zhang, 2008; statistical algorithms. A SGeMS plug-in is composed of two
Journel, 2007). The user friendliness of SGeMS comes in part elements:
1. a shared library (‘.so’ file on Linux, ‘.dll’ file on Windows),
1. Environmental Earth System Science, Stanford University, Yang and and
Yamazaki Environment and Energy Building 4215, 473 Via Ortega,
Room 353, Stanford University, Stanford CA 94305, USA. 2. a description of the graphical interface used to prompt for
Email: [email protected] parameters (‘.ui’ file).
A BOUCHER
The shared library is obtained by compiling the source code, CZ(h). The covariance matrices CII, CpI and Cpp, respectively the
and the user interface is created with the Qt Designer software covariance between the conditioning nodes, between the group to
(distributed with SGeMS). Writing a SGeMS plug-in requires be simulated and the conditioning nodes and between the nodes
moderate knowledge of the C++ language – a college level of the group. These covariance matrices must be computed to
introductory course is sufficient. SGeMS uses object-oriented build the LU system of equation with zI as the data vector
programming and template programming and modifying an contained in Ω i − 1 and zp the vector of the group of nodes to be
existing algorithm is the easiest and most efficient strategy to simulated:
create new algorithms.
⎡z I ⎤ ⎡L II 0 ⎤⎡wI ⎤
The first step in writing a plug-in is to define a new class z = Lw = ⎢ ⎥ = ⎢ ⎥⎢ ⎥
derived from the predefined class Geostat_algo and to redefine ⎣z p ⎦ ⎣L pl L pp ⎦⎣w p ⎦
the three virtual functions:
where:
1. initialise,
⎡C II C Tpl ⎤ ⎡L II 0 ⎤⎡ LTII LTpl ⎤
2. execute, and ⎢ ⎥=⎢ ⎥⎢ ⎥
3. name. ⎢⎣C p I C pp ⎥⎦ ⎣L pl L pp ⎦⎢⎣0 LTpp ⎥⎦
The purpose of these functions follows: and:

• Initialize() – Parse the input from the user and initialise the wp is a vector of independent realisation from a standard
parameters used by the execute function. These parameters Gaussian distribution
are usually stored as member variables of the algorithm class.
The simulated group of nodes zp is then:
• Execute() – Runs the algorithm using the parameters
recorded with the initialise() function. z p = C plC −1
II z 1 + L pp w p
• Name() – Return the name of the algorithm. A normal score transform may be needed before and after
• Graphical user interface (‘.ui’ file) – Create a user interface simulation to transform the data to and from a Gaussian
to record the user parameters – it is automatically loaded distribution.
with the plug-in. The g-sgsim algorithm can be summarised as follows:
Any algorithm derived from Geostat_algo can be seamlessly • perform a normal score transform (if necessary), and
integrated into the SGeMS platform. A simple example would be • initialise a random path visiting each group of pixels once.
a wrapper function that takes the user inputs, reformats it into a
For each group of pixels along the random path:
parameter file and calls an external program, for instance an
existing FORTRAN code. Once compiled, the plug-in (.dll or • get the neighbouring hard data and previously simulated
.so) should to be copied in the plugin/Geostat/ folder where nodes,
SGeMS is installed. When SGeMS is launched, all shared • build the LU system with the neighbourhood data,
libraries in that folder are loaded and they appear in the list of
available algorithms. • simulate the group by multiplying the L matrix with a vector
of independent standardised Gaussian realisations,
PLUG-IN EXAMPLE – THE GENERALISED • set the simulated points as data, and
SEQUENTIAL GAUSSIAN SIMULATION • back-transform the simulated data from a Gaussian
distribution to its original distribution (if necessary).
The generalised sequential Gaussian Simulation (g-sgsim) From an algorithmic perspective the differences between
(Dimitrakopoulos and Luo, 2004) is a generalisation of the sgsim and g-sgsim are:
sequential Gaussian simulation (sgsim) where instead of
simulating one point at a time with a simple kriging system, a • the simulation path visits discontinuous groups of nodes
group of points are simultaneously simulated with a LU instead of individual nodes,
decomposition algorithm (Davis, 1987). The g-sgsim algorithm • the drawing is done with the LU decomposition instead of
is faster than the traditional sgsim algorithm, an important aspect simple kriging, and
when simulated realisations of deposits are large, for example, • additional user input required about the size and geometry of
tens of millions of nodes (Benndorf and Dimitrakopoulos, 2007). the group.
The multivariate probability distribution function can be
written in term of Np -points group decomposition:
WRITING THE G-SGSIM PLUGIN
Ν1 N1 + N p = N2
This section shows only the relevant parts of the C++ code that is
f ( u1 ,..., uN ; z1 ,..., zN Ω 0 ) = Π f ( ui ; zi Ω i − 1 ). Π f ( ui ; zi Ω i − 1 ).
i=1 i = N1 + 1 modified from the sgsim algorithm. In addition to the initialise
Nv− 1 + N p = Nv
and execute functions, a new sequential simulation function that
is specifically written to handle group simulation is presented.
Π f ( ui ; zi Ωi−1 )
i = Nv− 1 + 1 The lines starting with the ‘//’ characters indicate comments and
where: are not part of the actual code. The ‘…’ characters indicate code
that is not presented here for brevity.
Ω0 is the set containing the observed (sampled) data
Ω i − 1 is the set of the conditioning data when simulating the ith
Function initialize
point
The function initialize() is a slightly modified version of the one
The domain D to be simulated is partitioned into Nv groups of in the sgsim algorithm. In the Gsgsim version, a group
Np points. For example, each group may identify a mining block. simulation path and a group_cursor are initialised instead of a
In the case where Np = 1, g-sgsim is the same algorithm as sgsim. regular simulation path. The group_cursor extracts the pixels
Let Z(u) be a Gaussian RF with mean mz and covariance model belonging to any given group along the simulation path.
MINING GEOSTATISTICS WITH THE OPEN SOURCE SGEMS SOFTWARE AND THE GSTL LIBRARY
Gsgsim::initialize (const Parameters_handler* parameters, harddata_property_ =

Error_messages_handler* errors ) { geostat_utils::gaussian_transform_property(
... harddata_property_, target_cdf_, harddata_grid_ );
// Instantiate the simulation path with the user-specified group }
// (nx,ny,nz) // Instantiate the object lu_sim to perform LU simulation with
path_ = new Gsgs_sim_path(simul_grid_,GsTLTriplet[nx,ny,nz]); // covariance covar_, the random number generator gen on the
// grid simul_grid_. The code to generate LU simulation is already
// Instantiate an object that retrieves all the points within a // available from the the GsTL library.
//group of size (nx,ny,nz)
LU_Simulator lu_sim(covar_,gen,simul_grid_);
group_cursor_ = new Group_sequential_cursor(simul_grid_, // Loop over all realisations
GsTLTriplet[nx,ny,nz]);
for( int nreal = 0; nreal < nb_of_realisations_ ; nreal ++ ) {
// Instantiate and parameterise a neighbourhood class with user- // Get a new group random path
// specified ranges and angles. path_->init_random_path();
neighborhood_ = SmartPtr<Neighborhood>( // Perform the simulation from the first group (path_->begin()) to
simul_grid_->neighborhood( ranges, angles, & // the last group (path_->end()) using the group_cursor_ to get the
covar_ ) ); // points within a group and lu_sim to draw the point realisations
... int status = group_sequential_simulation(
} path_->begin(), path_->end(),
*(neighborhood_.raw_ptr()), marginal,
Function execute progress_notifier.raw_ptr(),group_cursor_,
lu_sim );
The execute function generates nreal simulations specified by the
// Back-transform the realisation contained in property prop from a
user and parameterised by the initialise function. The function
// Gaussian distribution to the target_cdf_
first performs the normal score transformation of the data and
initialises the LU simulator with the correct covariance. Then for if( use_target_hist_ ) {
each simulation, it initialises a new random path and executes the cdf_transform( prop->begin(), prop->end(),
sequential simulation on the groups. The realisations are then marginal, target_cdf_ );
back-transformed from a Gaussian distribution to their original }
distribution (if requested by the user). ...
}
Gsgsim::execute(GsTL_projects*) {
...
// Perform normal score transform (if necessary) using the non- Function group_sequential_simulation
//parametric distribution object target_cdf_ (defined in GsTL) as This function implements the sequential simulation on groups
//distribution for the data. instead of points. For each group along the random path (from
if(use_target_hist_) { begin to end), it finds the points within that group with the object
FIG 1 - SGeMS interface with the algorithm panel (left) and the visualisation panel (right).
A BOUCHER
group_cursor, finds the neighboring data to that group of points

with the object neighbors and finally passes the unknown pixels
and the known values (both data and previously simulated
realisations) to the LU simulator.
int group_sequential_simulation( GeovalueIterator begin,
GeovalueIterator end, Neighborhood& neighbors,
const MarginalCdf& marginal, Notifier* notifier,
GroupCursor* group_cursor, GroupSimulator
group_simulator )
{
int bad = 0;
int loc = 0;
//Loop over all groups of nodes in the grid; the ordering has
//been already determined by the object path_. Geovalue is an
object that contains all the information about a point on a grid
such
//as its value and its location.
for(; begin != end; begin++, loc++) {
std::vector<Geovalue> group_gval;
group_cursor->get_group_gval(group_gval,*begin);
if( group_gval.empty() ) continue;
// Find the neighborhood data
neighbors.find_neighbors( * begin );
// Perform the simulation of the group of points by passing the
// unknown values (contained in group_gval.begin() to
// group_gval.end()), the known neighbouring data (contained in
// neighbors.begin() to neighbors.end()) and the marginal
// distribution (in this case a Gaussian with zero mean and unit
// variance) to group_simulator (in this case lu_sim).
// group_simulator returns zero when succesfull.
bad += group_simulator(group_gval.begin(),group_gval.end(),
neighbors.begin(),neighbors.end(),marginal);
// Send a signal to the progress bar and check if the user had
// cancelled the simulation in progress.
if( !notifier->notify() ) return -1;
}
return bad;
} FIG 2 - G-sgsim user interface.
User interface even for commercial products. By providing an interface to

Python, SGeMS allows users to automate repetitive tasks while
The user interface, shown in Figure 2, is derived by adding a taking advantage of the rich variety of Python libraries. Python is
widget to input the size of the group to be simulated. The QT an object-oriented language with a long list of specialised
program Designer is used to modify the user interface – no actual libraries to perform mathematical operations (eg numerical
coding is involved for the interface changes in most cases. integration and optimisation), image analysis, data visualisation
or even create graphical user-interfaces.
Examples on a iron ore deposit Some examples of common tasks performed with Python:
The Gsgsim plug-in described above is used to simulate an iron • Importing data – read data from any format and import them
deposit in Western Australia – see Boucher and Dimitrakopoulos into SGeMS.
(2007) for more details on the deposit. The Gsgsim user interface
and the simulation parameters are shown in Figure 2. The drill
• Exporting data – extract data from SGeMS and use Python to
export them into a desired format.
hole locations and Fe content are shown in Figure 3 and Fe
simulations generated with the plug-in with groups of 50 points • Data processing – get data from SGeMS and modify data
(5 × 5 × 2) are displayed in Figure 4. A total of 9 973 720 points values such as truncating high/low values, doing indicator or
were simulated. Figure 5 shows a 3D rendering of the low iron logarithm transform, upscaling, computing local varying
values, which correlate with high clay content. means, etc.
• Sensitivity analysis – repeated calls of an algorithm with
Automating tasks with Python different parameters.
It is often necessary to repeat tasks, eg to study the sensitivity of • Calling external programs – make SGeMS interact with an
an algorithm to different input parameters. Such sensitivity external program such as a mine schedule optimiser. This can
analysis would be tedious if the user had to manually update the be a great time saver for workflows that need multiple calls
parameters and re-run the algorithm. Embedded Python script in to programs.
SGeMS enables to automate such tasks. The Python language
The script must then do four steps:
(http://www.python.org/) is an interpretative language, ie a
program that does not require compilation to run. Python is 1. export data from SGeMS (simulation, estimation, etc) in
distributed under an open source license that makes it free to use, the right format;
FIG 3 - Drill hole locations and Fe values.
FIG 4 - One realisation from the Gsgsim algorithm for an iron ore deposit.
2. write parameter file(s) for the external program;

3. run the external program(s) using Python; and
4. import, process and display the results in SGeMS.
Example – cross-validation
Cross-validation is a common tool for choosing the right
parameters for a given geostatistical algorithm. The fitness of a set
of parameters is measured by estimating locations that are known
but removed from the data set, and computing the errors between
these estimates and the true values. A zero error indicates un- FIG 5 - Three-dimensional rendering of the low simulated Fe
biasedness and smaller squared errors indicates a better accuracy. values indicating high clay content.
To perform cross-validation, the Fe drill hole data set presented
earlier is first partitioned into K non-overlapping set. Each finally automatically displayed on a graph as shown in Figure 6.
partition set is then successively removed from the data set and The ‘#’ character indicates comments.
estimated using the neighbouring data belonging to the K-1
remaining set. Each sampled location is estimated only once. The
performance of the algorithm is assessed by analysing the errors
kCrossValidation() function
between the true and the estimated values.
The Python script below performs cross-validation by looking # Cross-validation function
at errors in function of the variogram types (spherical, ex- # k : Numbers of partitions of the data
ponential or Gaussian) and the variogram range in the horizontal # grid : Name of the sgems grid
plane. The kCrossValidation function takes as input k, the # hd_prop : Name of the property containing the data
number of partitions for the data set, the name of the grid and # v_ranges: Vector containing the variogram ranges
property of the data set, the variogram ranges in x, y and z # v_type : String with the variogram type
direction and the type of variogram to consider for ordinary # Seed: Initialise the random number generator (Optional)
kriging. The main program gathers the parameter for each call of
the kCrossValidation function, considering variogram ranges def kCrossValidation(k, grid, hd_prop, v_ranges, v_type, seed =
varying from 20 to 500 per increment of 20. The process is 111) :
repeated for a spherical, exponential and Gaussian type of errors = 0.
variogram. The errors are collected in a vector, saved in a file and errors2 = 0.
A BOUCHER
FIG 6 - Output from the Python script. Cross-validation errors (top) and squared errors (bottom) in function of the variogram range and type.
# Get the data from SGeMS kriging_values = scipy.array(sgems.get_property(grid,’__jk__’) )

data = scipy.array( sgems.get_property(grid, hd_prop) ) sgems.execute(‘DeleteObjectProperties
data_k = data.copy() dh::__jk__::__jk___krig_var’)
# Initialise the random number generator and randomise the # Compute the sum of errors and squared errors
indices of the data errors += scipy.sum(kriging_values[id] - data[id] )
scipy.random.seed( seed ) errors2 += scipy.sum( (kriging_values[id] -data[id])*
index = scipy.random.permutation( scipy.arange( len(data)) ) (kriging_values[id] - data[id]) )
# For each of the k partitions # Delete the temporary property and return an error vector
for i in range(k): sgems.execute(‘DeleteObjectProperties dh::__jk_hd__’)
# Get the indices for the k-th data partition (randomly selected) return [errors,errors2]
id = (index >= len(data)/k*(i)) == (index <
len(data)/k*(i+1)) Main program
# Set the k-th partition to uninformed and send the data to SGeMS
data_k[id] = -9966699 import sgems # functions to interact with SGeMS
sgems.set_property(grid,’__jk_hd__’, data_k.tolist()) import scipy # scientific computation module
# Replace back the uninformed data with the actual values import pickle # function to save complex data structures on file
data_k[id] = data[id] import pylab # plotting module
# Execute ordinary kriging with the specified variogram ranges # Create an empty dictionary to store the cross-validation errors
and variogram type errors = {}
sgems.execute(‘RunGeostatAlgorithm errors2 = {}
kriging::/GeostatParamUtils/XML::<parameters> # Vectors containing the variogram values to be tested
v_range = range(20,501,20)
<algorithm
v_ranges_str = [str(i) for i in v_range]
name=“kriging”/> <Grid_Name value=‘‘ ‘+grid+’ ” /> v_types = [‘Spherical’, ‘Exponential’, ‘Gaussian’]
<Property_Name value=“__jk__” /> <Kriging_Type # Loop all variogram types
type=“Ordinary Kriging (OK)” > <parameters /> for t in v_types :
</Kriging_Type>
errors[t] = []
<do_block_kriging value=“0”/> <Hard_Data
grid=‘‘ ‘+grid+’ ” errors2[t] = []
# Loop variogram ranges
property= “__jk_hd__” /> <Search_Ellipsoid
for r in v_ranges_str :
value=“500 500 10
# Call the cross-validation function with specific ranges (r) and
0 0 0” /> <Min_Conditioning_Data value=“0” />
variogram type (t)
<Max_Conditioning_Data value=“25” />
<Variogram results = kCrossValidation(50,’dh’,’FE_capped’,
[r,r,’15’] , t)
nugget=“0.2”structures_count=“1” > <structure_1
# Store the errors in the dictionaries
contribution=“0.8” type=‘‘ ‘+v_type+’ ” >
<ranges max=‘‘ ‘+v_ranges[0]+’ ’’ medium= ‘‘ errors[t].append(results[0])
‘+v_ranges[1]+’ ” errors2[t].append(results[1])
min=‘‘ ‘+v_ranges[2]+’ ” /> <angles x=“0” y=“0” # Save the results in a file
z=“0” /> ...Code omitted for brevity
</structure_1> </Variogram> </parameters>’) # Plot the errors and the squared errors in function of the range
# Extract and delete the kriging estimates from SGeMS ...Code omitted for brevity
This Python cross-validation code can easily be modified to Deutsch, C and Journel, A, 1998. GSLIB Geostatistical Software Library
analyse the sensitivity to the nugget effect, anisotropy, search and User’s Guide, second edition (Oxford University Press).
neighbourhood or the relevance of using nested variogram Dimitrakopoulos, R and Luo, X, 2004. Generalised sequential Gaussian
models. simulation on group size v and screen-effect approximations for large
field simulations, Mathematical Geology, 36:567-591.
Journel, A G, 2007. Roadblocks to the evaluation of ore reserves the
CONCLUSIONS simulation overpass and putting more geology into numerical
models of deposits, in Orebody Modelling and Strategic Mine
SGeMS offers many opportunities for mining applications. The Planning, second edition (ed R Dimitrakopoulos), pp 29-33 (The
extensive and open source geostatistical library provides a wide Australasian Institute of Mining and Metallurgy: Melbourne).
range of geostatistical algorithms and a sound framework for Osterholt, V and Dimitrakopoulos, R, 2007. Simulation of orebody
adding new algorithms. It offers opportunities to design algo- geology with multiple-point geostatistics – Application at Yandi
rithms specific for a mining site – the parameters and the user Channel iron ore deposit, WA and implications for resource
interface can be tailored for the needs of a particular application, uncertainty, in Orebody Modelling and Strategic Mine
such as taking account of topographical trends, non-conventional Planning, second edition (ed R Dimitrakopoulos), pp 51-59 (The
data transforms or even directly link to a company database. Some Australasian Institute of Mining and Metallurgy: Melbourne).
parameters can be set to values chosen by experts for rapid on-site Remy, N, 2004a. S GeMs: The Stanford geostatistical modelling
use by general practitioners. The addition of Python scripts software: A tool for new algorithms development in Geostatistics
Banff 2004 (eds: O Leuangthong and C V Deutsch), pp 865-872
provides a powerful way to streamline routine tasks and interface (Springer: Dordrecht).
with external programs. Finally, both academia and industry
Remy, N, 2004b. Algorithmic and software methods for a better
would benefit from a large repository of plug-ins for mining integration of the geological information into numerical models,
purposes where new algorithms are routinely contributed by users. PhD dissertation, Stanford University.
Remy, N, Boucher, A and Wu, J, 2009. Applied Geostatistics with
REFERENCES SGeMS: A User’s Guide (Cambridge University Press).
Remy, N, Schtuka, A, Levy, B and Caers, J, 2002. GsTL: The
Benndorf, J and Dimitrakopoulos, R, 2007. New efficient methods for geostatistical template library in C++, Computers and Geosciences,
conditional simulation of large orebodies, in Orebody Modelling and 28:971-979.
pp 103-110 (The Australasian Institute of Mining and Metallurgy: Strebelle, S, 2002. Conditional simulation of complex geological
Melbourne). structures using multiple-point statistics, Mathematical Geology,
34:1-21.
Boucher, A and Dimitrakopoulos, R, 2007. A new efficient joint
simulation framework and application in a multivariable deposit, in Wu, J, Boucher, A and Zhang, T, 2008. SGeMS code for pattern
Orebody Modelling and Strategic Mine Planning, second edition simulation of continuous and categorical variables: FILTERSIM,
(ed: R Dimitrakopoulos), pp 303-312 (The Australasian Institute of Computers and Geosciences, 34:1863-1876.
Mining and Metallurgy: Melbourne).
Davis, M, 1987. Production of conditional simulations via the LU
decomposition of the covariance matrix, Mathematical Geology,
19:91-98.
HOME
Conditional Simulation Methods to Determine Optimum Drill

Hole Spacing
C Boyle1
ABSTRACT 2001; Chapman, 2003; Journel and Kyriakidis, 2004). Kriging

estimation variance is only the conditional variance in the case
Resource evaluation drilling is usually on a regular pattern; the pattern
spacing should be optimised to maximise profit from mining, accounting
of a Gaussian random function with strict stationarity, such that
for the cost of drilling and the value of additional information from simple kriging can be used as the estimation technique. The
increased density of drilling, and also to reduce risk in mined ore tonnes distribution of estimation errors is likely to be skewed if the
and grades to an acceptable level. Conditional simulation methods for variable estimated has a negatively or positively skewed
determining optimum drilling spacing are more powerful than traditional distribution, or where the point or block being estimated is in a
methods, and simulation methods can take into account local variability high or low grade area with respect to the deposit average. It is not
in grade. Mining profit functions can consider profit and density of possible to sum kriging variances, as kriging variances of adjacent
evaluation drilling so that profit can be maximised. A conditional blocks are spatially correlated and not independent. This makes it
simulation method, here termed the simulation-estimation method, impossible to compare kriging variance estimation errors for
consists of the following: making conditional simulations of the orebody various selective mining unit sizes.
using the evaluation drilling; sampling these realisations to produce sets Conditional simulation takes into account local grade
of artificial drill holes; estimating resources using the artificial drill holes;
variability, such that areas with higher values can be modelled
and comparing estimated resources with the realisation from which the
respective artificial drill holes were taken. The effectiveness of the with higher variances between different simulations
simulation-estimation method is demonstrated in a case study with the (Dimitrakopoulos, Godoy and Chou, 2010, in this volume).
exhaustive Walker Lake dataset in which predicted results, using a Conditional simulation methods can determine appropriate drill
sampled dataset, are similar to those from the exhaustive dataset. The hole spacing in areas exhibiting the proportional effect (local
simulation-estimation method is applied to a mining case study of a variance correlated with local mean) and heteroscedasticity
Western Australian bedded iron ore deposit and shows that estimation (fluctuating values of local mean and variance), rather than
errors are higher for the impurities SiO2, Al2O3, and P, rather than Fe. A homoscedasticity (constant mean and variance), as assumed in the
mining profit function suggests that the preferred resource evaluation drill use of kriging variance.
hole spacing to maximise mining profit is a spacing of 40 × 40 m, which
has a higher expected profit than spacings of 10 × 10 m, 20 × 20 m and 80 ×
80 m. Charts of drill hole spacing versus expected average error allow a METHODOLOGY
spacing to be selected that has the maximum allowable level of average
error. Conditional simulation methods can also be used for quantitative Relative absolute error
resource classification that considers mining production rates and
selectivity, as well as for highlighting periods of high risk in mine plans Simulation allows various sampling patterns and optimum drill
due to high uncertainty in ore grade. hole spacing to be assessed quantitatively by calculating various
measures like relative absolute error, amount of misclassification
INTRODUCTION of blocks in a resource model and the economic consequences of
estimation errors. With closer drill hole spacing there are more
Resource evaluation drilling of ore deposits to allow estimation conditioning data and the various simulations will be increasingly
of tonnes and grades is usually carried out on either a regular alike. In the end case, if there could be complete sampling of the
square or rectangular grid pattern. Resource evaluation data is deposit, then all simulations would be the same. One approach,
needed for pit design, long-term scheduling, determining termed herein the simulation-estimation method, is documented
blending requirements and stockpiling studies. As spacing of by Li et al (2007) and is as follows in the eight steps below:
drilling decreases, there is a rapid increase in cost for drilling,
sampling, assaying, and interpretation, while uncertainty of 1. Make a point simulation of the deposit at a fine scale
grades and orebody location decrease. Optimum spacing can be conditional to sampling and geological data. The size of the
quantitatively evaluated using the cost of collecting extra blocks must be at least as fine as the closest drill hole
sampling versus economic benefits of this additional information spacing to be tested later.
so that mining profit is maximised. Resource evaluation drilling
2. Sample simulations at a specified drill hole spacing to
is usually not the final grade control method for allocation of ore
or waste categories. Final grade control is often done just prior to generate a set of artificial drill holes.
mining, using additional sampling data collected from dedicated 3. Build a model of the deposit based on the set of artificial
reverse circulation (RC) percussion drilling or from blast holes. drill holes, by means of the usual technique for resource
Historically, ordinary kriging estimation variance was used to estimation of the deposit (often ordinary block kriging).
determine accuracy of various drill hole spacings (David, 1988; The model of the deposit is generated at the same block
Journel and Huijbregts, 1978). Once enough evaluation drilling size as for normal resource estimates.
data has been collected to produce a reasonable semivariogram,
4. Reblock the points in the fine initial simulation to the same
various patterns of drill holes can be designed and dummy
support as blocks in the resource model in step 3 above.
ordinary kriging can be conducted based on the patterns of drill
holes. This can be done without actually drilling the holes, as This reblocking is used to represent the same support size
kriging variance only depends on the configuration of data values used in reality, and must include a sufficiently large number
and not on the magnitude of the values themselves. However, there of nodes, so that the simulated models used represent the
are a number of problems with the use of kriging variance to actual variability of the blocks.
determine the estimation error (Goovaerts, 1997; Schofield, 5. Calculate the relative absolute error of each block in the
estimated deposit developed in step 3 by comparing the
block grade to reality from step 4. This can be done on a
1. Email: [email protected] block by block basis, or a number of blocks can be
C BOYLE
combined into volumes that represent mining over different The number of holes of a smaller drilling grid increases by a
periods, such as months or years. Relative absolute error is multiplicative factor relative to the number of holes on a larger
calculated by the formula: grid size in the following relationship:
E ij = Vsij - Ve j / Vsij (i = 1,..., n; j = 1,..., m) (1) ( )( )

fdrill hole = 1 + Xlarge / Xsmall . 1 + y large / y small / 4 (3)
where: where:
Eij is the relative absolute error of block j estimated using
fdrill hole is a multiplicative factor giving the relative change in
simulated deposit i
number of drill holes with change in drill hole spacing
Vsij is the reblocked simulated value of block j
xlarge is the x axis spacing of the larger drill grid
Vej is the estimated value of block j
xsmall is the x axis spacing of the smaller drill grid
n is the number of simulated deposits
ylarge is the y axis spacing of the larger drill grid
m is the number of blocks
ysmall is the y axis spacing of the smaller drill grid
6. A relatively high number of simulations (eg 30 to 50) are As an example, halving the grid spacing from 80 m to 40 m in
made at a fine scale, repeating step 1 above. both x and y axes will increase the number of holes by a factor
7. Repeat steps 2 to 5, using a variety of drill hole spacing fdrill hole of 9/4:
covering the range of possible drill hole spacing options.
fdrill hole = (1 + 80/40)(1 + 80/40)/4 = 9/4
8. Make charts of the average relative errors for different drill
hole spacings, both globally and for local areas. Maps can also If 400 drill holes are needed to cover the extent of an orebody
be produced showing relative errors for different regions. with an 80 m grid, then 900 drill holes are needed to cover the
Relative absolute error can be calculated for blocks of same area with a 40 m grid.
selective mining unit size or on larger volumes after combining
adjacent blocks in a model. The larger volumes may represent TESTING THE METHOD WITH AN EXHAUSTIVELY
mining production for quarters or years, as determined from KNOWN DEPOSIT
long-term schedules of future mining. Grades of larger volumes
should be closer to true grades than grades of smaller volumes to
true grades, due to averaging of grades over the larger volumes. Data and statistics
The exhaustive Walker Lake elevation dataset consists of 78 000
Profit function measurements of a variable V from a regular 1 × 1 m grid that
extends 260 m east-west and 300 m north-south (Isaaks and
Drilling and sampling costs, as well as the potential benefits for Srivastava, 1989). The exhaustive data set is sampled on a 20 m
mine planning, can be calculated for various drill hole spacing square grid representing 13 lines (east-west) and 14 lines
patterns. Conditional simulation allows distributions of profits for (north-south) of drilling data from a mineral resource deposit at a
resource model blocks to be determined and the spacing with feasibility stage of evaluation (Figure 1). The sampled drilling
highest potential profit selected. Boucher, Dimitrakopoulos, and dataset is used with the simulation-estimation method to
Vargas-Guzman (2004) calculate profit per tonne mined for determine optimum drill hole spacing. Six square drill hole
various patterns of infill grade control drilling in a nickel cobalt spacings are used:
deposit and their approach is modified here for use in drilling of 1. 5 × 5 m,
an iron ore deposit at a resource evaluation stage, rather than at the
grade control stage. 2. 10 × 10 m,
Economic effects of block misclassification can be determined 3. 20 × 20 m,
using costs of sampling and drilling, and the profit from sale of
4. 30 × 30 m,
iron ore. US dollars are used throughout this study and all unit
costs and revenues are approximate for the general Australian 5. 40 × 40 m, and
iron ore industry only, rather than for any particular mine or 6. 50 × 50 m.
company. The main costs are taken to be for drilling, logging,
sampling, and interpretation – as well as mining, crushing and A spacing of 5 × 5 m refers to 5 mE × 5 mN and is for two
transportation from mine to port (by rail or truck). Costs of iron dimensions. Results are considered for 20 × 20 m squares,
ore shipping and later transport to steel mill are usually paid for representing blocks in a resource model. The 20 × 20 m areas are
by the customer; these costs are not included in this study. reblocked into 60 × 60 m, 100 × 100 m and 260 × 300 m areas,
representing mining production for various time periods. The
The cost of producing a block at location u in category k, is:
Walker Lake study considers grade uncertainty for various areas,
as is required in the mining industry for stockpiling and blending
C kTotal (u) = C kDrilling + C kCrushing + C kTransportation (2) studies to manage product or crusher feed variability. The
proportions above or below a cut off grade are not considered for
where: the Walker Lake study.
The statistics and shape of the sample data set distribution
C kTotal is total cost (Figure 2) closely resemble the distribution of the exhaustive
Drilling data set (Figure 3). The sample data set has a positively skewed
C k is drilling, logging, sampling, and interpretation cost distribution with a moderately high coefficient of variation.
C kMining is mining cost Directional semivariograms are calculated for V with a lag
Crushing distance of 20 m and directions aligned with the sample lines.
C is crushing cost
k The semivariogram model has a nugget of 25 000 (40 per cent),
Transportation
C k is transportation cost from mine to port one spherical structure with a sill of 38 000 (60 per cent), and
CONDITIONAL SIMULATION METHODS TO DETERMINE OPTIMUM DRILL HOLE SPACING
Simulation-estimation method
The simulation-estimation method uses the steps summarised
above. Fifty conditional simulations are made by the turning
bands method with 400 bands (Chiles and Delfiner, 1999; Boyle,
2007). The search neighbourhood is circular with radius 100 m, a
minimum of three and maximum of 16 samples, and no use of an
octant search (as drilling is not clustered). Simulations in normal
scores space are back transformed to data space. Point
simulations have no artificial striping, and honour means,
variances and semivariograms of sampling data. Point
simulations are reblocked to 20 × 20 m squares and 20 × 20 m
squares are reblocked to 60 × 60, 100 × 100 and 260 × 300 m
squares. Each simulation is sampled at six grid spacings (5 × 5,
10 × 10, 20 × 20, 30 × 30, 40 × 40 and 50 × 50 m) to produce six
sets of artificial drill holes. Ordinary block kriging estimates
20 × 20 m areas, for a total of 300 kriging estimates, that are
reblocked to four sizes (20 × 20, 60 × 60, 100 × 100 and
260 × 300 m). Relative absolute error for each area is calculated
FIG 1 - Walker Lake sample locations. by Equation 1 above and the simulation-estimation method has
50 reality datasets. Relative absolute errors for each of the four
area sizes and six spacings are shown in Figure 4.
True relative absolute error

Results from the simulation-estimation method are compared
with true results from the exhaustive data. True results are those
produced by ordinary kriging (as in a resource estimate) of
various spacings of drill holes extracted from the exhaustive
values, with the assumption of no sampling error. Estimates
made using true values are compared with the exhaustive data
reblocked to the same sizes (20 × 20, 60 × 60, 100 × 100 and
260 × 300 m). Averages of true relative absolute errors for each
of the four area sizes and six spacings are shown in Figure 4.
Results
Values of relative absolute error generally increase with
increasing drill hole spacing (Figure 4), though there are a few
exceptions due to whether specific samples are selected, for
FIG 2 - Histogram of variable V for samples at Walker Lake. example with 100 × 100 m areas between 40 m and 50 m drill
hole spacing. This is more pronounced for wider drill hole
spacings, as the influence of each sample is higher. The charted
lines for the simulation-estimation method plot within a
reasonable distance to lines of the true values, so the
simulation-estimation method generally performs acceptably in
FIG 3 - Histogram of variable V for exhaustive data at Walker Lake.
ranges of 40 m east and 55 m north. Samples at 20 × 20 m,

representing evaluation drilling, are transformed to normal
scores by a Hermite polynomials method. Directional
semivariograms are calculated for normal scores of V and the
semivariogram model has a nugget of 0.2, one spherical FIG 4 - Summary of relative absolute errors of the simulation-
structure with a sill of 0.8, and ranges of 40 m east and 60 m estimation method and relative absolute errors of true values for V
north. over four areas and using six drill hole spacings at Walker Lake.
C BOYLE
this study. However, there is overestimation of average relative internal waste BIF that have separate interpreted wireframes, and
absolute error with closer spaced drilling patterns, for 10 m and samples of these BIF pods are not included in statistics and
20 m spacing in 20 × 20 m areas and for 10 m to 30 m spacing in simulation of the mineralised domain. Ideally, the full space of
260 × 300 m areas. At 30 to 50 m spacing in 20 × 20 m areas, the uncertainty of a mineral deposit is assessed by various geological
simulation-estimation method has similar results to true values. interpretations or domain simulations, for example as by
Drill hole spacing required for a defined relative absolute error Osterholt and Dimitrakopoulos (2007), but for this study only
depends on the area to be assessed. True values for an average deterministic wireframes are used and instead grade variability
relative absolute error of less than ten per cent for areas of within the ore domain is examined.
20 × 20 m require a 5 m square drilling spacing, whereas if the Fe is negatively skewed with a low coefficient of variation
areas are 260 × 300 m, then only 50 m drilling spacing is compared with Al2O3, SiO2 and P (Figures 5 to 8). Fe has strong
sufficient for a similar relative absolute error. This indicates a negative correlations with Al2O3 and with SiO2. Direct
requirement for early discussions between geologists and mining semivariograms of Al2O3, Fe, P and SiO2 are modelled in data
engineers with regards to planning drilling, expected mining space for use in estimation.
selectivity and production rates, so that drilling spacing is
optimised for evaluation. If production rates are high and
selectivity is low in a bulk mining operation, then early drilling
expenditure can be reduced by using a wider spacing, whereas
selective mining will require a tighter drill hole spacing if the
mine is to achieve expected production and cash flow.
In this Walker Lake study, involving one variable, one domain
and six possible drill hole spacings, the simulation-estimation
method uses 50 simulations, then 300 kriging estimations. There
is also the associated reblocking and calculations of relative
absolute error. Efficient software with a fast simulation algorithm
and scripting capabilities are prerequisites for handling the large
amount of processing in a drill hole spacing study by simulation
in a reasonable time. In summary, the simulation-estimation
method has performed reasonably in the Walker Lake study; the
same method is used in the following iron ore case study.
APPLICATION – THE WELD RANGE IRON ORE

DEPOSIT STUDY
FIG 5 - Histogram of Al2O3 in drill hole composites at Weld Range.
Data and statistics
The simulation-estimation method of determining drill hole
spacing for evaluation drilling is applied to the W14 orebody of
the Weld Range iron ore deposit (Neehoff et al, 1998). Weld
Range is located 65 km northwest of Cue, and 720 km
north-northeast of Perth in the Murchison area of Western
Australia. Mineralisation consists of a 4 km strike length of
massive goethite-haematite lenses, up to 80 m thick, dipping at
70° and located between a footwall of Archaean banded iron
formation (BIF) and hanging wall of dolerite. Canga (surface
re-cemented iron ore) is also present in addition to the massive
goethite-haematite lenses. Open file data provided by the
Western Australian Department of Industry and Resources from
the 1997 - 1998 drilling program at Weld Range are used for this
case study (Neehoff et al, 1998). This public domain drilling data
consists of 124 reverse circulation (RC) holes (WWRC01-
WRRC123) and 22 diamond holes (WR01-WR28), assayed for
Fe, SiO2, Al2O3, P, LOI, Mn, S, TiO2, Ca, MgO and K – with
density determination by downhole geophysical logging. Drilling
samples are composited to 2 m for equal support. These data are FIG 6 - Histogram of Fe in drill hole composites at Weld Range.
an example of what is available at a feasibility stage of resource
evaluation. Additional drilling as well as a feasibility study has
taken place at Weld Range since the 1997 - 1998 drilling Simulation-estimation method
program, but none of that company domain data has been used in
this study. The area modelled in this study is between 47 480 - Normal scores for Al2O3, Fe, P and SiO2 are created by a
50 520 mE (local grid). Hermite polynomials method. Direct and cross semivariograms
Optimum drill hole spacing depends on the most variable of of normal scores are modelled by a linear model of
the important chemical elements, rather than on less important or coregionalisation (Goovaerts, 1997) for use in cosimulation. A
more spatially continuous elements. For iron ore at Weld Range, spacing below of 10 × 10 × 2 m, for example, means 10 mE ×
the important four chemical variables have been identified as Fe, 10 mN × 2 mRL – three dimensions are considered in the Weld
SiO2, Al2O3 and P – only these are examined in this case study. Range study.
The majority of drilling samples have assays for all four
The eight steps for the Weld Range study are:
variables. Sectional drill hole interpretation defines a single main
mineralised domain of greater than 50 per cent Fe, hanging wall 1. Fifty point cosimulations of grid spacing 10 × 10 × 2 m are
dolerite and footwall BIF. The mineralised domain has pods of generated and checked.
8. Profits are calculated for each block when mine planning is

carried out using different evaluation drill hole spacings
and the spacing that maximises profit is determined, taking
into account the costs of drilling, mining, crushing and
transportation.
Results
Average relative absolute errors vary widely for Al2O3, Fe, P and
SiO2 for quarterly and yearly production periods (Figures 9 to
16). For each variable, the average errors of quarterly periods are
slightly higher than yearly periods for the same variable.
Averages of quarterly periods and yearly periods for the four drill
hole spacings are summarised in Figure 17. Fe errors are an order
of magnitude lower than for the other variables, due to the high
Fe concentration and lower coefficient of variation. Quarter 20
has a higher error than most other quarters, but even for quarter
20 on 80 × 80 × 2 m drill hole spacing, the average relative
FIG 7 - Histogram of P in drill hole composites at Weld Range. absolute error of Fe is less than one per cent (Figure 10). Fe may
be less important than the contaminants in deciding drill hole
spacing due to the higher variability of contaminants. The
highest average relative absolute errors of contaminants are for
SiO2, followed by Al2O3 and then P. SiO2 relative absolute errors
average less than two per cent at 10 × 10 × 2 m spacing for both
quarterly and yearly averages, and increase monotonically to
greater than four per cent at 80 × 80 × 2 m spacing for both
quarterly and yearly periods (Figure 17). Various drill hole
spacings for different variables can be evaluated for their
respective average expected errors, and selection made of the
widest spacing with the required maximum allowable average
error. SiO2 may be the critical variable in determining drill hole
spacing due to its highest average relative absolute error and
importance as an impurity to steel mills. If average SiO2 relative
absolute error must be less than three per cent, then a spacing of
40 × 40 × 2 or 50 × 50 × 2 m is needed for both quarterly and
yearly periods of production (Figure 17).
Average relative absolute errors for quarters of each variable
differ considerably (Figures 9 to 12) due to mining areas with
different variability (heteroscedasticity). Due to correlation
FIG 8 - Histogram of SiO2 in drill hole composites at Weld Range. between variables, periods with high Fe average errors often also
have high average errors for Al2O3, P and SiO2. Quarters 12 and
2. Simulations are sampled at spacings of 10 × 10 × 2, 20 × 20
× 2, 40 × 40 × 2 and 80 × 80 × 2 m, representing 50 sets of
artificial drill holes.
3. Ordinary block kriging estimations are made using each
drill hole spacing for each sampled simulation, for a total
of 200 estimations. Ordinary block kriging is a common
estimation method for an iron ore resource model.
Estimations are for blocks of size 40 × 40 × 10 m.
4. The 50 initial simulations are reblocked to the same size as
for the blocks in the kriging estimations in step 3 above
(40 × 40 × 10 m). Reblocked simulations represent 50
representations taken to be reality.
5. Relative absolute error of each block in the estimated deposit
developed from a particular simulation in step 3 is calculated
by comparing estimated block grades to reality for that
simulation from step 4 (Equation 1). Errors are calculated
over the W14 orebody for production periods of quarters
(20 quarters of about 3 Mt) and years (five years of around
12 Mt). Tonnage calculations assume an in situ bulk density
of 3.2 kg/m³.
6. Divide the orebody into quarterly and yearly periods of
production as in a conceptual mine plan with 10 m mining
FIG 9 - Relative absolute errors for Al2O3 over quarterly periods
benches.
and estimated using four drill hole spacings at Weld Range.
7. Charts of relative absolute errors are created for quarterly Drill hole spacings are 10 × 10 × 2, 20 × 20 × 2, 40 × 40 × 2 and
and yearly production periods. 80 × 80 × 2 m.
C BOYLE
20 have high average errors predicted and are thus high risk,
whereas quarters 2 and 3 have low average errors predicted
(Figures 9 to 12). Average error charts for years mainly agree
with those for quarters – year 5 has higher than usual risk in
predicting ore grades (Figures 13 to 16). If expected grade
prediction errors are too high, then pre-emptive action can be
taken, for example by additional drilling or adding stockpiling
and blending capacity. Alternatively, multiple mining areas or
pits could be scheduled for periods of highest risk, allowing more
blending options and choice of mining blocks available to
overcome the higher expected variability. If mine plans rely on a
few mining areas of high average error, then the plans are high
risk. Expected financial penalties due to variability of impurities
in ore can be evaluated and an assessment of the value of
additional drilling can be made. Methods using conditional
FIG 12 - Relative absolute errors for SiO2 over quarterly periods

and estimated using four drill hole spacings at Weld Range.
FIG 10 - Relative absolute errors for Fe over quarterly periods and

estimated using four drill hole spacings at Weld Range.
FIG 13 - Relative absolute errors for Al2O3 over yearly periods and
estimated using four drill hole spacings at Weld Range. Drill hole
spacings are 10 × 10 × 2, 20 × 20 × 2, 40 × 40 × 2 and 80 × 80 × 2 m.
simulations can show areas of higher risk better than methods

using only drill hole spacing or kriging variances, as the latter
assume equal estimation error if sampling density is equal,
whereas errors are likely to be greater in areas with higher or
lower values of chemical variables than average.
Classification has been made into ore or waste for resource
model blocks, for each of 50 simulations (each taken to be
reality) and 50 sets of drill holes taken from sampling each
simulation at four drill hole spacings. Using each simulation as
reality, the proportion of misallocation of blocks can be
FIG 11 - Relative absolute errors for P over quarterly periods and determined. In an iron ore deposit, grades of impurities can be as
estimated using four drill hole spacings at Weld Range. important to steel mill customers as the grade of Fe. Three
FIG 14 - Relative absolute errors for Fe over yearly periods and

estimated using four drill hole spacings at Weld Range. FIG 16 - Relative absolute errors for SiO2 over yearly periods and
estimated using four drill hole spacings at Weld Range.
FIG 15 - Relative absolute errors for P over yearly periods and FIG 17 - Summary of average relative absolute errors of the
estimated using four drill hole spacings at Weld Range. simulation-estimation method over quarterly and yearly periods for
four drill hole spacings at Weld Range.
chemical parameters are used for block classification in the Weld Proportions of ore and waste blocks are determined for
Range study to define high grade ore for direct shipping to blast simulations (taken to be reality) and kriged estimates from
furnaces:
sets of drill holes are taken from each simulation. Block
1. Al2O3 <5 per cent, misclassification decreases with decreasing drilling spacing
2. Fe >57 per cent, and (Figure 18).
Prices paid for iron ore are based on metric tonne units (MTU)
3. SiO2 <10 per cent. of Fe metal in $US/MTU, either for fines or lump iron ore. This
These high grade chemical parameters are generalised for this study has been simplified to not consider lump and fines
study and are not from any particular mine. P has not been percentages and uses a constant price of $US0.90/MTU
considered in block classification, as the mean grade of the (EconStats, 2008), with no moisture adjustment. The revenue per
samples is fairly low at 0.07 per cent and P levels should be able tonne of ore is calculated as:
to be controlled by blending during mining. Instead, Al2O3, Fe,
and SiO2 are more significant when deciding drill hole spacing. Rore = Fe % x RMTU (4)
C BOYLE
where: where:
Rore is the revenue per tonne of ore Pore-ore is profit for actual ore modelled as ore
Fe % is the Fe grade of ore
Pwaste-waste = -$5 - CDrilling (6)
RMTU is the revenue per metric tonne unit of Fe metal
where:
Pwaste-waste is profit for actual waste modelled as waste (always
negative)
Pore-waste = (Fe % × $0.90 - $5 - $1 - $3 - CDrilling) × 0.8 (7)
where:
Pore-waste is profit for actual ore modelled as waste
Pwaste-ore = -$5 - CDrilling (8)
where:
Pwaste-ore is profit for actual waste modelled as ore
Ptotal = ∑ i = 1 (Pore-ore + Pwaste-waste + Pore-waste + Pwaste-ore)

j
(9)
where:
Ptotal is total profit for all blocks 1 to j depending on the
category of each block j
FIG 18 - Weld Range proportion of ore and waste in mineralised Note that Pwaste-waste is assumed equal to Pwaste-ore,
ore domain. Legend format is <actual>_<modelled>, for example
in the legend ore_waste is for actual ore modelled as waste.
Paverage = Ptotal ÷ j (10)
In the case of 60 per cent Fe and $US0.90/MTU, the revenue where:

per tonne of ore is 60 per cent × $0.90/MTU = $54/tonne. Paverage is the average profit of all mined blocks 1 to j (ore and
Production costs per tonne in $US have been assumed to be: waste)
• mining $5, Paverage can be easier to use than Ptotal when selecting a drill hole
• crushing, spacing and is calculated for each drill hole spacing (Figure 19).
• screening $1, and
• trucking or railing to port $3.
Drilling, sampling, assaying, and interpretation costs, C kDrilling ,
are assumed to be $100 per metre for reverse circulation (RC)
drilling and on four possible drill hole spacings have costs per
tonne for ore and waste of 10 × 10 m at $1.084, 20 × 20 m at
$0.271, 40 × 40 m at $0.038 and 80 × 80 m at $0.017. There are
four categories, k, for mining blocks:
1. actual ore modelled as ore,
2. actual ore modelled as waste,
3. actual waste modelled as ore, and
4. actual waste modelled as waste.
It is assumed that later grade control will efficiently determine
the correct allocation of blocks (the information effect is not
taken into account in this study), mining blocks of 40 mE, 40 mN
and 10 mRL represent the selective mining unit, and that drilling
Drilling
costs C k are the same for ore and waste. Hence one cost
CDrilling suffices for all k (category). A discount factor of 0.8 is
applied to true ore estimated to be waste, with the assumption FIG 19 - Profit by drill hole spacing at Weld Range.
that most ore estimated to be waste lies inside the pit shell and
will be eventually located by grade control sampling, but that
timely mining of this ore will be delayed and mine planning will Even though a 10 × 10 m spacing will result in the most
be suboptimal. accurate estimation with the highest proportions of correctly
Profit calculations for each category, Pk, calculated with assigned blocks (Figure 18), the far higher cost of drilling for a
Equation 2 above are: 10 × 10 m spacing means that a 10 × 10 m spacing does not
maximise profit. For a spacing of 80 × 80 m, misallocation of ore
and waste blocks is too high to compensate for the lower cost of
Pore-ore = Fe % × $0.90 - $5 - $1 - $3 - CDrilling (5) drilling. Instead, a spacing of 40 × 40 m has the highest profit,
FIG 20 - Cumulative density function of relative absolute errors for 100 × 100 m areas at Walker Lake.
$US33.0/tonne, of four considered drill hole spacings and so and Indicated mean the relative absolute errors between kriged
40 × 40 m is the preferred drill hole spacing at the resource grades or tonnages relative to simulated reality over the volumes
evaluation stage for the Weld Range iron ore deposit. representing quarterly or yearly production are less than
15 per cent for 90 per cent of the volumes. Charting cumulative
distribution function (CDF) of block relative absolute errors
RESOURCE CLASSIFICATION allows determination of error at a 90 per cent confidence level.
The JORC code for ore reserves and resources is prepared by the This approach is applied for the Walker Lake study, assuming a
Joint Ore Reserves Committee of the Australasian Institute of quarterly production area is 100 × 100 m and a yearly production
Mining and Metallurgy, Australian Institute of Geoscientists and area is 260 × 300 m. For 90 per cent of 100 × 100 m areas, the
Minerals Council of Australia and requests a statement of the relative absolute error is below 34 per cent for a 5 m drill hole
relative error or confidence in the resource estimate, if possible by spacing and below 57 per cent for a 50 m drill hole spacing
quantitative statistical or geostatistical procedures (Joint Ore (Figure 20), so no drill hole spacings considered allows a
Reserves Committee, 1994). This can be done by conditional Measured classification. For 90 per cent of 260 × 300 m areas,
simulation approaches as in the above studies of drill hole spacing. relative absolute error is below eight per cent for a 5 m drill hole
Though there are currently no generally accepted quantitative spacing, below 12 per cent for a 30 m drill hole spacing and
criteria for resource classification, Parker (2004) proposes that a below 22 per cent for a 50 m drill hole spacing (Figure 21). The
Measured resource classification is appropriate for volumes widest spacing with a relative absolute error below 15 per cent at
of quarterly production where there is less than ±15 per cent 90 per cent confidence is the 30 m drill hole spacing, which
error with 90 per cent confidence, with confirmed continuity therefore allows an Indicated classification.
of mineralisation. An Indicated resource classification is The Weld Range study considers relative absolute errors in
appropriate for volumes of yearly production where there is less grade within a deterministic ore-BIF boundary. This study could
than ±15 per cent error with 90 per cent confidence, with be extended to simulate ore-BIF boundaries (Osterholt and
assumed continuity of mineralisation. Error applies to both Dimitrakopoulos, 2007), followed by grade simulation within the
tonnes and grades of ore. Conditional simulation can determine simulated boundaries, to produce a more complete model of the
expected error for different drill hole spacings and allow for a space of uncertainty. This would allow a quantitative
competent person to decide resource classification. Applied to determination of tonnes and grade of ore that could be used for
the simulation-estimation method, these definitions for Measured resource classification at Weld Range.
C BOYLE
FIG 21 - Cumulative density function of relative absolute errors for 260 × 300 m areas at Weld Range.
CONCLUSIONS REFERENCES
The simulation-estimation method performs reasonably for the Boucher, A, Dimitrakopoulos, R and Vargas-Guzman, A, 2005. Joint
Walker Lake case study, as shown by comparison of average simulations, optimal drillhole spacing and the role of the stockpile, in
relative errors of block grades calculated by ordinary kriging Geostatistics Banff 2004, Proceedings Seventh International
Geostatistics Congress, Canada, pp 35-44 (Springer: Netherlands).
using a sampled subset of simulations, when compared with
those of true kriging results expected from sampling of Boyle, C, 2007. Comparison of multivariate conditional simulation
methods at the Yandicoogina iron ore deposit, in Proceedings Iron
exhaustive data. Errors predicted decrease as the size of the Ore 2007, pp 57-66 (The Australasian Institute of Mining and
production areas assessed increase – larger selective mining Metallurgy: Melbourne).
units and higher production rates allow a wider evaluation drill Chapman, E, 2003. An investigation using conditional simulation to assess
hole spacing to be used. the relationship between the level of confidence in the estimated
A profit function developed to determine optimum drill hole resource and the drilling grid at the Warramboo pisolite iron ore
spacing for resource evaluation and used in the Weld Range deposit, Robe River, Masters of Science in Mining Geology thesis
iron ore case study, shows that maximum profit is achieved for (unpublished), Camborne School of Mines, University of Exeter,
a 40 × 40 m drill hole spacing. Of the major chemical variables Cornwall, United Kingdom.
for iron ore, average expected relative absolute error is greatest Chiles, J P and Delfiner P, 1999. Geostatistics, Modelling Spatial
Uncertainty, 695 p (John Wiley and Sons: New York).
to least for SiO2, Al2O3, P and Fe. For each variable, average
errors for quarterly production periods are slightly greater than David, M, 1988. Handbook of Applied Advanced Geostatistical Ore
Reserve Estimation, 217 p (Elsevier: The Netherlands).
for yearly production periods for the same variable. Charts of
Dimitrakopoulos, R, Godoy, M and Chou, C L, 2010. Resource/reserve
drill hole spacing versus average relative absolute error show a
classification with integrated geometric and local grade variability
suitable spacing for the required level of estimation error. measures, in Advances in Orebody Modelling and Strategic Mine
Resource classification can be determined quantitatively for Planning I (ed: R Dimitrakopoulos), pp 215-222 (The Australasian
expected production rates and selectivity. Mine plans can be Institute of Mining and Metallurgy: Melbourne).
examined using conditional simulation methods to show EconStats, 2008. Iron ore prices and producers [online], EconStats.
periods of high risk and action, like extra drilling or plan Available from: <http://www.econstats.com/rt_ironore.htm>
changes, taken if risk is unacceptable. [Accessed: 10 November 2008].
Goovaerts, P, 1997. Geostatistics for Natural Resources Evaluation, 483 p Neehoff, B, Miller, A, Dumlao, F and Llorca, J, 1998. An Feng Kingstream
(Oxford University Press: New York, USA). Steel Weld Range Project TR70/3902 and E20/176, Exploration report
Isaaks, E and Srivastava, R, 1989. An Introduction to Applied for the period September 1997 - November 1998, Department of
Geostatistics, 561 p (Oxford University Press: New York). Industry and Resources Western Australia, open file report A56957.
Joint Ore Reserves Committee, 2004. The JORC Code, Australasian Osterholt, V and Dimitrakopoulos, R, 2007. Simulation of orebody
Code for Reporting of Exploration Results, Mineral Resources and geology with multiple-point geostatistics – Application at Yandi
Ore Reserves. Available from: <http://www.jorc.org> [Accessed 10 Channel iron ore deposit, WA and implications for resource
November 2008]. uncertainty, in Orebody Modelling and Strategic Mine Planning,
Journel, A and Huijbregts, C, 1978. Mining Geostatistics, 600 p second edition (ed: R Dimitrakopoulos), pp 51-59 (The Australasian
(Academic Press: London). Institute of Mining and Metallurgy: Melbourne).
Journel, A and Kyriakidis, P, 2004. Evaluation of Mineral Reserves a Parker, H, 2004. Resource and Reserve Audits, Geostatistical Association
Simulation Approach, 232 p (Oxford University Press: New York). of Australia talk, Perth, Western Australia, 10 May.
Schofield, N, 2001. Determining optimal drilling densities for near mine
Li, S, Dimitrakopoulos, R, Scott, J and Dunn, D, 2007. Quantification of
resources, in Mineral Resources and Ore Reserve Estimation – The
geological uncertainty and risk using stochastic simulation and
AusIMM Guide to Good Practice, (ed: A Edwards), pp 293-298 (The
applications in the coal mining industry, in Orebody Modelling and
Australasian Institute of Mining and Metallurgy: Melbourne).
Melbourne).
HOME
Transformation Methods for Multivariate Geostatistical

Simulation – Minimum/Maximum Autocorrelation Factors and
Alternating Columns Diagonal Centres
E M Bandarian1, U A Mueller2, J Fereira3 and S Richardson4
ABSTRACT can be described as a nugget model together with an intrinsic

To speed up multivariate geostatistical simulation it is common to
spatial component (Tran, Murphy and Glacken, 2006). The use of
transform the set of attributes into spatially uncorrelated factors that can a 2SLMC to derive the transformation matrix is largely of
be simulated independently. The main method in recent years has been theoretical interest and while there are versions of MAF that
minimum/maximum autocorrelation factors, either based on the work directly with the model (Bandarian, 2008), in most
coefficient matrices of a two structure linear model of coregionalisation implementations it is common to work with experimental matrices
(LMC) or on a pair of experimental covariance matrices. In both cases directly as proposed by Desbarats and Dimitrakopoulos (2000),
there is an underlying assumption that the covariance structure of the data thus cutting down on the effort spent in finding a reasonably well
set can be adequately modelled using a two structure LMC. We consider fitting 2SLMC. As a trade-off however time needs to be invested
an extension that removes the restriction imposed by this assumption by in choosing the lag spacing at which to calculate the experimental
using the experimental matrices for a larger set of lags. The method relies
semi-variogram matrix. Whichever approach is chosen, be it
on an iterative algorithm that approximately diagonalises a set of
symmetric matrices, and is referred to as the Alternating Columns-
theoretical or experimental, the factors calculated from the input
Diagonal Centres method. We use the Jura data set to evaluate the extent data are only approximately spatially orthogonal.
to which factors obtained from each method are spatially decorrelated Because of the restrictions in the MAF-method, a more general
and to assess the effect of the transformation method on the simulated approach such as joint approximate diagonalisation (JAD) can be
attributes. considered. JAD consists of finding a transformation matrix that
will approximately diagonalise a set of matrices. These techniques
INTRODUCTION have been developed for image-filtering and blind noise separation
(see for example Yeredor (2002) and references therein). JAD
The simulation of one or more attributes in a multivariate data set algorithms for approximate diagonalisation of symmetric matrices
is enhanced through the exploitation of the relationships between can be broadly split into two categories: those that derive an
them. However, multivariate modelling and simulation can be orthogonal diagonalising matrix like that proposed by Manton
time consuming and computationally inefficient. Some of the (2005) and those that weaken the orthogonality constraint to non-
shortcomings of multivariate methods can be avoided by singularity of the diagonaliser. The algorithm we will consider is
transforming the spatially correlated set of attributes into a set of of the latter type and is referred to as the Alternating Columns
spatially uncorrelated factors. These factors can then be modelled Diagonal Centres (ACDC) method (Yeredor, 2000, 2002). For
and simulated independently using univariate geostatistical geostatistical applications the set to be jointly approximately
methods. The oldest method for obtaining uncorrelated factors is diagonalised consists of experimental semi- variogram matrices.
principal component analysis (Wackernagel, 2003). While this The transformation matrix derived from this method will
technique is easy to implement, the factors are uncorrelated for all approximately spatially decorrelate the given attributes, in the
distances other than zero only in the case of intrinsic correlation. sense that the cross correlations between distinct attributes are
In recent years the method of minimum/maximum autocorrelation approximately equal to zero.
factors (MAF) has received much attention as the factors are In the standard implementation, MAF is applied to normal
derived so that they are approximately spatially uncorrelated for scores data and the resultant factors are subsequently simulated.
all distances (Desbarats and Dimitrakopoulos, 2000; Boucher There is no theoretical requirement to apply MAF to standard
and Dimitrakopoulos, 2007). This method assumes that the data normal data (Bandarian, 2008) and we will be taking the
to be modelled are multivariate normal and so in most cases the approach to calculating the required transformation matrices
first step is to normalise the data. While the method of direct from standardised raw data instead. The reason for this approach
minimum/maximum autocorrelation factors (DMAF) (Bandarian, is partly the observation that the application of MAF to normal
Bloom and Mueller, 2008) removes the requirement of normality scores can result in transformed data that are no longer normally
of the data and so simplifies the MAF procedure, one still requires distributed, thus resulting in the need for a further normal scores
the assumption of a two structure linear model of coregionalisation transform if a Gaussian simulation algorithm is to be used.
(2SLMC) in order to determine the transformation coefficients. In what follows, we present a brief review of the MAF and
Extensions of MAF to more than two structures are not possible ACDC methods and illustrate them using a subset of the Jura
(Vargas-Guzman and Dimitrakopoulos, 2003), unless the model data set (Goovaerts, 1997). The transformations are calculated
from the correlation matrix and semi-variogram matrices of
standardised data. The factors resulting from each of the methods
1. Mathematics Teacher, Cyril Jackson Senior Campus, 53 Reid Street, are transformed to normal scores in order to implement sequential
Bassendean WA 6054. Email: [email protected] Gaussian simulation (SGS). The normal scores simulations are
2. Associate Professor, School of Engineering, Edith Cowan University,
back-transformed to factors, then to the original data space. The
270 Joondalup Drive, Joondalup WA 6027. resultant simulations are assessed for reproduction of the spatial
Email: [email protected] and statistical characteristics of the sample.
3. Honours Student, School of Engineering, Edith Cowan University,
270 Joondalup Drive, Joondalup WA 6027. PROBLEM DATA
Email: [email protected] A four variable subset of the Jura data set (Goovaerts, 1997)
4. Lecturer, School of Engineering, Edith Cowan University, consisting of Cd, Co, Cr and Ni is used to illustrate each
270 Joondalup Drive, Joondalup WA 6027. transformation method and the resultant simulations. Location
Email: [email protected] maps (see Figure 1) for the sample data show no obvious trends
E M BANDARIAN et al
FIG 1 - Location maps for cadmium, cobalt, chromium and nickel. The colour scales are based on the deciles of the respective sample data.
TABLE 1
Summary statistics for the input data, for each attribute there are 259 values.
Variable Mean Variance Skewness Minimum LQ Median UQ Maximum

Cd 1.31 0.84 1.50 0.14 0.63 1.07 1.72 5.13
Co 9.30 12.79 -0.18 1.55 6.52 9.76 12.00 17.72
Cr 35.07 120.07 0.29 8.72 27.44 34.84 42.32 67.60
Ni 19.73 67.78 0.16 4.20 13.72 20.56 25.44 53.20
in the spatial distribution of the data. The sample data have quite TABLE 2
disparate means and variances (see Table 1) and of the four
Correlation matrix for sample data.
metals, only Chromium is approximately normally distributed
(Bandarian, 2008). The sample correlation coefficients indicate Variable Cd Co Cr Ni
varying strengths of linear relationships, with those between
Cd 1.00 0.25 0.61 0.49
Cd-Cr, Co-Ni and Cr-Ni having the stronger correlations (Table 2).
For stability of implementation the data are first standardised by Co 0.25 1.00 0.45 0.75
subtracting the sample mean from each value then dividing by Cr 0.61 0.45 1.00 0.69
the sample standard deviation.
Ni 0.49 0.75 0.69 1.00
MULTIVARIATE TRANSFORMATION METHODS

AND SIMULATION step 2 – independently model and simulate factors, and
T
Let Z(u)=[Z1(u), Z2(u),...ZK(u)] be a vector of K second-order step 3 – back transform simulated factors to simulated attributes.
stationary random functions over a study region A, let Step 1 includes the calculation of the transformation matrix
z(u)=[z1(u ), z2(u ),...zK(u )]T, α=1,...n, be the corresponding from the data, in the case of MAF this requires in the first
isotopic sample vectors and let B denote the correlation matrix of instance a decision what lag spacing to choose for the calculation
the random function. of the experimental semi-variogram matrix, for ACDC a decision
The general procedure for applying transformation methods needs to be made as to how many semi-variogram matrices
for multivariate simulation is: are to be included in the computation. In addition in step 1
step 1 – transform the data to approximately decorrelated an assessment needs to be made whether or not the resultant
factors, factors are sufficiently well decorrelated. Assuming the spatial
TRANSFORMATION METHODS FOR MULTIVARIATE GEOSTATISTICAL SIMULATION
decorrelation is satisfactory, then step 2 is only concerned with experimental semi-variogram Γ$ (h1 ) matrix calculated at lag
the univariate simulation of the factors. Either a direct simulation spacing h1 can be used to approximate Γ(h) at h1. In this case the
algorithm, such as the sequential simulation proposed by Soares MAF transformation matrix is derived from the generalised
(2001) or a standard Gaussian algorithm, such as Sequential eigenvalue problem reformulated in terms of the correlation
Gaussian Simulation can be used. In the latter case a Gaussian matrix B and the experimental semi-variogram matrix Γ$ (h1 ) at
anamorphosis will need to be carried out to compute normal lag h1 where h1 < a (a is the maximum range of the
scores for the factors. For both methods, Step 3 consists merely semi-variogram). Regardless of the suitability of the 2SLMC the
of the backtransformation of the simulated factors via matrix transformation matrix X simultaneously diagonalises Γ$ (h1 ) and
multiplication. The resultant realisations are still standardised B yielding factors which are orthogonal at lag spacings zero and
and so shifting and scaling to restore the appropriate mean and h1 only.
variance respectively concludes the workflow. A crucial assumption in the construction of the MAF factors is
the requirement that the covariance structure of the underlying
Minimum/maximum autocorrelation factors random function is fully characterised by a 2SLMC. This
assumption is restrictive in practice and extensions to more
For MAF it is assumed that the semi-variogram function, (h), of general models have been discussed in the literature (see for
the multivariate random function Z(u) can be modelled by a two example the discussion in Vargas-Guzman and Dimitrakopoulos
structure linear model of coregionalisation (h)=B1g1(h)+B2g2(h), (2003). However, in general, for three covariance structures or
where the symmetric coregionalisation matrices B1 and B2 more there exists a matrix X that diagonalises the variogram only
contain the sills of the permissible semi-variogram models g1(h) when all but (at most) one of the coefficient matrices in the LMC
and g2(h), and B=B1+B2 is the correlation matrix. Since B1 and are proportional to one another (Tran, Murphy and Glacken,
B2 are symmetric (therefore diagonalisable) and B2 is positive 2006; Bandarian, 2008).
definite (has positive eigenvalues), then B1 and B2=B-B1 may be
diagonalised simultaneously by congruence (that is, there exists a
non-singular matrix X, not necessarily orthogonal, such that Alternating Columns Diagonal Centres
S=XTAX) and the diagonalising matrix is the solution to the
symmetric definite generalised eigenvalue problem (Datta, 1995): The Alternating Columns Diagonal Centres (ACDC) method
(Yeredor, 2002, 2004) iteratively determines a matrix X that
B1X=(B-B1)X diagonalises a set of J symmetric K × matrices {B1,B2,...,BJ}
The diagonal matrix in the equation above is the matrix of by minimising:
generalised eigenvalues and the matrix X is the matrix of J
Ψ(XD1,D2,...,DJ) = ∑ w j B j − XD j XT
2
generalised eigenvectors. ,
F
j=1
The equation may be rewritten as:
B1X=BX 1
where ⋅ F denotes the Frobenius norm, wj > 0, j = 1,..., J are
(optional) weights and {D1,D2,...,DJ} is a set of diagonal
where: Λ1=Λ(I+Λ) -1
matrices. The algorithm consists of alternating phases, the AC
The generalised eigenvalue problem may be converted into a phase, where the objective function is minimised with respect to
standard eigenvalue problem with a symmetric matrix on the left
the columns of X while the set {D1,D2,...,DJ} is kept fixed and
hand side of the equation using the Cholesky factorisation of the
matrix B where B=LLT and L is a non-singular lower triangular the DC phase, where the objective is minimised with respect to
matrix. If we put G=L-1B1(LT)-1 and Y=LTX, then GY=YΛ1. {D1,D2,...,DJ} while X is kept fixed.
The matrix G is symmetric by construction and so Given {B1,B2,...,BJ} the algorithm is initialised by specifying
orthogonally diagonalisable. The matrix X obeys either an initial set of diagonal matrices {D1,D2,...,DJ} or an
XTBX=YTL-1B(LT)-1Y=YTL-1LLT(LT)-1Y=YTY=I and the estimate of the diagonalising matrix X. The initial phase in the
column vectors in X are B-orthonormal, that is XTBX=I. iteration is dependent on the specification made: if an initial
Furthermore, XTB1X=YTL-1B1(LT)-1Y=YTGY=Λ1. guess for the diagonalising matrix is made, then the algorithm
starts with a DC phase, otherwise the starting point is an AC
Thus the non-singular matrix X, whose columns are the phase. In the absence of either specification the diagonalising
generalised eigenvectors xk,k=1,...,K, simultaneously diagonalises matrix is set to the identity matrix and the algorithm is initialised
the matrices B1 and B by congruence (Bandarian and Mueller, with a DC phase. Thus the standard iterations are as follows:
2008).
given {B1,B2,...,BJ},
The MAF factors are derived by putting FMAF(u)=XTZ(u). $ = I.
put X
Since the matrix X is B-orthogonal the factor variances are
equal to one and the transformation ensures that the factors F(u)
are uncorrelated for all h ≥ 0 because: DC phase
Γ F(h) = X Γ(h)X
T
1. $ TX⊗X
Set G = [X $ TX]-1
T T
= X B1Xg1(h)+X (B-B1)Xg2(h) 2. For j = 1,..J
= Λ1g1(h)+(I-Λ1)g2(h) $ = diag(G diag(X
Set D $ TB X)).
$
j j
Hence Γ F(h) is diagonal for all h.
The MAF transformation thus diagonalises the semi-variogram AC phase
model of the attributes exactly. However, the associated
experimental semi-variograms of the MAF-factors are only For k = 1,...K
J
⎡ K
⎤
$ [ j] ⎢ $ T ⎥
approximately diagonalised. 1. Set P = ∑ w j λ k B j − ∑ λ n xnx n
[k ]
$ $
⎢ ⎥
Where the theoretical LMC is not known the experimental j=1
⎢⎣
n =1
⎥⎦
semi-variogram matrices Γ$ (⋅) are used to calculate the MAF
n ≠k
transformation coefficients. Assuming that the semi-variogram 2. Find the largest eigenvalue of P and an associated
function of Z(u) is fully characterised by a 2SLMC, an eigenvector of unit norm.
E M BANDARIAN et al
a μ Perfect spatial decorrelation occurs when ζ(h) = r(h) = 0 and

3. If μ < 0 set x$ k = 0 otherwise put x$ k = k.
κ h) = 1 for all lag vectors h. A set of factor semi-variogram
∑w ( )
J
λ$ [kj ]
2
j matrices may be considered to be nearly in diagonal form if κ(h)
j=1 ≥ 0.9 for all h lags (Xie, Myers and Long, 1995).
The algorithm alternates between the two phases and halts Global measures for the spatial decorrelation are the averages
once a prespecified tolerance has been reached. The algorithm is ζ, τ and K calculated over J lag spacings:
applied to a sequence of semi-variogram matrices, calculated at
1 J 1 J 1 J
increasing distance from the origin. There are no requirements
other than real symmetry of the matrices. If X denotes the
ζ= ∑
J j=1
ζ(h j ), τ = ∑ τ(h j ) and K = ∑ k(h j )
J j=1 J j=1
matrix that approximately diagonalises the given semi-variogram
matrices then the ACDC factors are derived putting Simulations and reproduction of sample
FACDC (u ) = X−1 Z(u ). Thus for both methods the factors are characteristics
derived using a linear transformation, however neither of the
matrices is orthogonal, as would be the case for PCA. The For each attribute, the suite of realisations is assessed for
transformations are both local transformations, in that they act on reproduction of the attribute (target) histogram, semi-variogram,
the attributes at each location, but the matrix coefficients are mean, variance and correlation coefficients. Qualitative analysis
constant across the entire study region. Provided the factor scores consists of visual inspection of the histogram and semi- variogram
are sufficiently spatially orthogonal the factors may be swarms of the realisations and boxplots of the realisation means,
independently modelled and simulated. In this paper, a standard variances and correlation coefficients about the target statistic.
Gaussian algorithm will be used for the factors. As a first step Quantitative measures for the reproduction of the target
the factors are transformed to normal scores (Gaussian statistics are the semi-variogram and histogram mean square
anamorphosis), then they are independently modelled and deviation (VMSD and HMSD respectively) which are calculated
simulated using Sequential Gaussian Simulation (100 for each realisation. The HMSD is given by:
realisations). The simulated attribute scores are retrieved via
1 P
Gaussian anamorphosis from normal scores to factor scores, then
from factor scores to standardised scores using the inverse MAF
HMSD = ∑ ( z p − z1p )2
P p= 1
and ACDC transformation matrices. Finally the sample means
and standard deviations are reincorporated. where:
zp is the pth percentile of the target cumulative distribution
Performance measures function (cdf)
zlp is the pth percentile of the lth simulated cdf
Spatial decorrelation
P denotes the total number of percentiles calculated
The quality of the spatial decorrelation of the factors will be The VMSD is given by:
assessed both graphically and numerically. For a graphical
assessment the cross semi-variogram for each factor pair will be 1 J
graphed in order to detect any remaining spatial correlation.
VMSD = ∑ ( γ$ j − γ$ lj )2
J j=1
Quantitative measures (Tercan, 1999) used to assess spatial
decorrelation are the absolute deviation from diagonality ζ(h), where γ$ j and γ$ lj are the target and simulated experimental
the relative deviation from diagonality τ(h) and the spatial semi-variogram values respectively calculated at the jth lag
diagonalisation efficiency K(h). spacing and J is the total number of lags.
The absolute deviation from diagonality, ζ(h) at lag h is
defined to be the sum of squares of the off-diagonal elements of RESULTS AND ANALYSIS
the factor experimental semi-variogam matrix at lag h:
Transformations
ζ(h) = ∑ k = 1 ∑ j ≠ k ( γ$ (h; k, j)) , h > 0
K K 2
For ACDC, the set of target matrices consists of thirteen
F
experimental semi-variogram matrices calculated at a lag spacing
where γ$ F (., k, j) denotes the experimental cross-semi-variogram of 0.2 km with a lag tolerance of 50 per cent. The weight vector
for the factors Fj and Fκ . was set to be w = [10 10 1 ... 1]T. The transformation matrix is:
The function τ(h) compares the absolute sum of off-diagonal ⎡ 0.841 − 0.048 − 0.289 − 0152
. ⎤
~ ⎢ 0.049
elements of the factor experimental semi-variogram matrix Γ F(h) 0.782 0.525 − 0.811⎥
with the sum of the absolute values of the diagonal elements X−ACDC
1
=⎢ ⎥
calculated at each lag h: ⎢−0.382 − 1.045 1.247 0.457⎥
⎢ 0132 − 0.966 1.939 ⎥⎦
⎣ . 0121
.
∑ ∑
K K
k=1 j≠k
γ$ F (h; k, j)
τ(h) = ,h > 0 The MAF transformation matrix, shown is obtained using the
∑ γ$
K
F (h; k, k) sample correlation matrix B and an experimental semi-variogram
matrix Γ$ (h)1 , calculated at h1 = 0.220 km:
k=2
Finally, the function K(h) compares the sum of squares of the ⎡−1.082 0.215 − 0.036 0.288⎤
off diagonal elements of the factor experimental semi-variogram ⎢−0.663
matrix Γ$ F(h) at lag h to the sum of squares of the off diagonal − 0.934 1.005 0.695 ⎥
XTMAF =⎢ ⎥
elements of the sample experimental semi-variogram matrix ⎢−0114
. 1176
. 1.062 − 1.493 ⎥
Γ$ z(h): ⎢−0152 − 0.305 − 0.955 ⎥⎦
⎣ . 0.485
∑ ∑
K K
k=1 j≠k
γ$ F (h; k, j))2 This spacing was chosen from a set of separation distances as it
κ(h) = 1 ,h > 0 provided the best overall decorrelation for MAF (Bandarian,
∑ ∑
K K
k=1 j≠k
( γ$ z (h; k, j)) 2
2008).
FIG 2 - Experimental cross semi-variograms of factors.
The cross semi-variograms for each factor set are displayed in of these measures for all lags (0.037 km to 2.407 km inclusive)
Figure 2. For ACDC the spatial decorrelation is excellent with and for the lower lags (0.037 km to 1.199 km inclusive) are
the majority of cross semi-variogram values satisfying shown in Table 4.
−015
. ≤ γ ik (h) ≤ 015
. . For this method there are only three
instances where the factor pairs still show some correlation. For TABLE 4
MAF, as expected, the spatial decorrelation is perfect at
Average (h), (h) and (h) for all lags and lower lags (0.037 km to
h1 = 0.220 km, while for other lag spacings the semi-variogram 1.199 km inclusive).
values are typically −0.2 ≤ γ ik (h) ≤ 0.2 (with the exception of
factors 2 and 4 at h1 = 1199
. km). These results are summarised in
All Lower
Table 3.
Plots of the measures of spatial decorrelation (ζ(h), (h) and ξ(h) τ(h) κ(h) ξ(h) τ(h) κ(h)
(h)) are displayed in Figure 3 and the corresponding averages ACDC 0.065 0.157 0.975 0.055 0.166 0.970
MAF 0.076 0.202 0.968 0.071 0.210 0.957
TABLE 3
Summary of factor pairs with remaining spatial correlation. The plots of ζ(h), (h) and (h) reflect the excellent spatial
decorrelation achieved by ACDC, which outperforms MAF for
Method Factor pairs Lag spacing γik
(km)
all lag spacings other than 0.220 km. The average diagon-
alisation efficiency for ACDC is 0.975 for all lags and 0.97 for
AC-DCAC-DC A3-A4 0.797 0.227 the lower lags, which is slightly higher than that achieved by
1.199 0.228 MAF where the values are 0.97 for all lags and 0.96 for the lower
2.407 −0.300 lags.
MAF D1-D3 0.595 −0.182 The factor scores from each method are transformed to normal
scores using a Gaussian anamorphosis with 50 Hermite
1.405 0.168
polynomials to approximate the factor histograms, independently
1.795 0.160 modelled then simulated using SGS. In order to avoid spurious
D2-D3 1.012 −0.163 correlations resulting from the use of the same random paths for
D2-D4 0.797 −0.182 each factor, the random number seeds were changed, so that
corresponding factors for the two methods used the same seed,
1.199 −0.221 but within the simulation of the factors the seeds were distinct.
D3-D4 2.196 0.178 The simulated attribute scores are retrieved via backtransformation
from normal scores to factor scores, then from factor scores to
standardised scores using the inverse ACDC and MAF
transformation matrices:
⎡ 1.258 0.416 0.277 0.207 ⎤
⎢−0189
. 1.024 − 0131
. 0.4445⎥
XACDC =⎢ ⎥
⎢ 0.215 0.862 0.660 0.222 ⎥
⎢ 0.033 0.585 ⎥⎦
⎣ 0.337 0.318
⎡−0.926 − 0.247 − 0.094 − 0.271⎤

⎢ 0.030 0.009 − 0.471 − 0.882 ⎥
(X )
−1
T
=⎢ ⎥
⎢−0.382 − 0.592 − 0.691 − 0164
MAF
. ⎥
⎢−0121 − 0.621 − 0.229 − 0.740 ⎥⎦
⎣ .
Finally the sample means and standard deviations are
FIG 3 - Plots of spatial decorrelation (ζ(h): n, (h): ∗ and (h): p). reincorporated.
E M BANDARIAN et al
FIG 4 - Mosaic maps of realisation 25 for Cd, Co, Cr and Ni.
Simulations
Mosaic maps for one of the realisations (picked at random)
generated for each metal are shown in Figure 4. The mosaic
maps are broadly similar. The spatial variability of the sample
data has been reproduced in the simulation.
Reproduction of the target statistics

The realisation cdf swarms overlaid with the corresponding
sample cdf for each attribute and transformation method indicate
that the attribute distributions have generally been reproduced
(see Figure 5). The most notable deviation from the target cdf
occurs for Co. For both methods the Co target cdf is poorly
reproduced on the interval from approximately 4 ppm to 6 ppm
with the realisation swarms being consistently lower than the
target cdf. This is more pronounced for ACDC than for MAF.
Boxplots of the HMSDs for each set of attribute realisations
(Figure 6) reflect the overall reproduction of the target cdfs. For
Cd and Co the HMSDs from MAF are slightly lower than those
for ACDC, for Ni and Cr, the situation is reversed. For Cd the
HMSDs are higher for ACDC than those for MAF, with the
former having 75 per cent of values less than 0.209 while the
latter has 75 per cent of values less than 0.160. The distribution
of HMSDs for Co is similar for both methods although those for
MAF are generally slightly lower than for ACDC. For Cr the
HMSDs for ACDC are typically lower than those for MAF with
the former having 75 per cent of values less than 2.007 while the
latter has 75 per cent of values less than 2.238. Similarly for Ni
where ACDC has 75 per cent of values less than 1.385 while FIG 5 - Histogram swarms.
MAF has 75 per cent of values greater than 1.507.
Figure 7 displays the boxplots of the realisation summary means are within ±5 per cent of the target for both methods. The
statistics. For both methods and all attributes the target means MSD of the means for each attribute (Table 5) are generally
have been reproduced with the majority of realisation means lower for MAF, with the exception of Co where ACDC is lower.
being within ±5 per cent of the target. For Cd, ACDC and MAF Reproduction of the target variances is similar for both
80 per cent and 84 per cent of realisation means respectively lie methods. In general the realisation variances are lower than the
within ±5 per cent of the target. For Co this increases to 98 per corresponding targets, however the majority of realisation
cent and 97 per cent respectively while for Ni the percentages are variances are within ±15 per cent of the target for both methods.
96 per cent and 97 per cent respectively. For Cr the realisation For ACDC Cd, Co, Cr and Ni have 84 per cent, 78 per cent
Reproduction of the target correlation coefficients (see Figure 8)

between attributes is similar for both methods. With the
exception of Cd-Co the targets have been reproduced with the
majority of realisation correlations being within ±10 per cent of
the target. For Cd-Cr, Co-Ni and Cr-Ni in excess of 96 per cent
of realisation correlations are within ±10 per cent of the target
for both methods. For Cd-Ni the percentages are 78 per cent and
75 per cent for ACDC and MAF respectively, while for Co-Cr
they are 84 per cent and 69 per cent respectively. For Cd-Co the
target correlation coefficient has not been reproduced by either
method, although to a lesser extent for ACDC than for MAF.
ACDC yields 27 per cent of realisation correlations within
±10 per cent of the target while for MAF this percentage is only
19 per cent. The MSD of the correlations for each attribute pair
FIG 6 - Boxplots of histogram mean square deviation. (Table 6) are similar for both methods, with ACDC being slightly
lower for Cd-Co, Cd-Ni and Co-Cr while for the remaining
attribute pairs MAF is slightly lower.
100 per cent and 62 per cent of realisation variances within
±15 per cent of the target respectively. For MAF these
percentages are 86 per cent, 84 per cent 98 per cent and 58 per TABLE 6
cent respectively. The MSD of the variances for each attribute Mean square deviation of correlation coefficients.
(Table 5) are generally lower for MAF, with the exception of Ni
where ACDC is lower. MSD Cd-Co Cd-Cr Cd-Ni Co-Cr Co-Ni Cr-Ni
Correlations AC-DC 0.065 0.029 0.040 0.035 0.040 0.026
TABLE 5 MAF 0.072 0.028 0.041 0.043 0.038 0.024
Mean square deviation of realisation means and variances.
MSD Cd Co Cr Ni
Reproduction of the target variograms
Means AC-DC 0.051 0.198 0.673 0.527 Experimental semi-variograms for the MAF and ACDC
MAF 0.045 0.223 0.515 0.462 simulations in Figure 9 indicate that both approaches have resulted
in adequate reproduction of the experimental sample semi-
Variances AC-DC 0.091 1.635 8.198 9.827
variograms. The variograms swarms are similar for both
MAF 0.088 1.485 7.439 9.974 transformation methods for each attribute and attribute pair. In
FIG 7 - Boxplots of realisation means and variances.
E M BANDARIAN et al
general the direct semi-variograms have been reproduced with

the overall shapes, ranges, nuggets and sills of the simulations
being similar to those of the target variograms. The most notable
exception to this is the nugget for Cd (both transformations)
where the target is somewhat lower than those of the simulations.
For the cross semi-variogram swarms, the simulations
obtained from both transformation methods are again similar. In
general the target cross semi-variograms have been reproduced
with respect to the overall shapes, nuggets and sills (except
Cd-Co where the target sill is underestimated in both cases).
With the exception of Cd-Cr the target ranges have been
overestimated. For Co-Ni, the cross semi-variogram swarm from
ACDC reproduces the target cross semi-variogram slightly better
FIG 8 - Boxplots of realisation correlation coefficients.
than the swarm resulting from MAF, in particular for lag
distances between 1 km and 1.5 km. However, for both methods
the cross semi-variograms have lower sills than the experimental
data. This is also featured in the correlation coefficients for this
attribute pair, for Co and Ni the correlation at lag 0 is lower than
that of the target.
The overall reproduction of the direct and cross semi-
variograms is reflected by the low values of the VMSDs for all
attributes and attribute pairs for both transformation methods
(Table 7 and Table 8). The majority of VMSDs are less than
0.15, with the exception of those for Co-Ni where the majority of
VMSDs are less than 0.2. The distribution of VMSD values is
generally similar for each transformation method, with results for
ACDC slightly better than for MAF.
For both ACDC and MAF, the transforms are constructed
from omnidirectional semi-variogram matrices. Of the data
under consideration Ni and Co exhibit anisotropy with direction
of greatest continuity approximately E-W for Ni and SW-NE
for Co. A comparison of the average experimental semi-
variograms of the simulations and the standardised Co data
indicates that both methods capture the anisotropy and that
there is almost no difference in the average directional
variograms calculated from the realisations (Figure 11).
Similarly for Ni, the average experimental semi-variograms
reproduce the E-W anisotropy (Figure 12). For Cr and Cd, the
average semi-variograms show the realisations to be isotropic.
Thus at the univariate level, the spatial features of the raw data
are broadly reproduced, even though the spatial decorrelation
methods were based on omnidirectional semi-variogram
matrices.
Similarly, an inspection of the directional cross-semivario-
grams for Co and Ni and the corresponding realisations reinforces
the observations from the omnidirectional cross-semivariogram
swams. There is good coincidence between the average
experimental cross variograms for both sets of realisations and the
corresponding data cross-variograms for shorter separation
distances, while for distances between 1 km and 1.5 km the sills in
FIG 9 - Experimental semi-variogram swarms (grey lines) and directions SW-NE and E-W do not reach the data sills, in direction
corresponding targets (•). SE-NW the averages fit well.
TABLE 7
Mean square deviation of semi-variogram values.
ACDC MAF
Cd Co Cr Ni Cd Co Cr Ni
min 0.098 0.043 0.049 0.064 0.096 0.04 0.061 0.071
lq 0.104 0.055 0.06 0.097 0.101 0.049 0.072 0.096
med 0.108 0.062 0.066 0.109 0.104 0.056 0.076 0.106
uq 0.111 0.07 0.073 0.122 0.108 0.064 0.08 0.117
max 0.118 0.09 0.095 0.15 0.115 0.087 0.091 0.149
mean 0.108 0.063 0.068 0.109 0.105 0.057 0.076 0.106
variance 0 0 0 0 0 0 0 0
TABLE 8
Mean square deviation of cross-semi-variogram values.
Cd-Co Cd-Co Cd-Cr Cd-Cr Cd-Ni Cd-Ni Co-Cr Co-Cr Co-Ni Co-Ni Cr-Ni Cr-Ni
(ACDC) (MAF) (ACDC) (MAF) (ACDC) (MAF) (ACDC) (MAF) (ACDC) (MAF) (ACDC) (MAF)
min 0.039 0.043 0.054 0.062 0.033 0.039 0.044 0.041 0.041 0.062 0.041 0.045
lq 0.069 0.074 0.067 0.072 0.079 0.08 0.068 0.075 0.121 0.112 0.077 0.091
med 0.083 0.088 0.08 0.081 0.101 0.102 0.085 0.093 0.141 0.137 0.098 0.108
uq 0.111 0.108 0.094 0.099 0.128 0.12 0.101 0.111 0.171 0.164 0.12 0.126
max 0.146 0.175 0.137 0.158 0.182 0.205 0.17 0.184 0.226 0.238 0.179 0.216
mean 0.088 0.093 0.083 0.086 0.104 0.103 0.088 0.097 0.143 0.138 0.098 0.109
variance 0.001 0.001 0 0 0.001 0.001 0.001 0.001 0.002 0.001 0.001 0.001
FIG 10 - Experimental cross semi-variogram swarms (grey lines) and corresponding targets (•).
CONCLUSION availability in commercial software, even though the need to use a

non-linear transformation may well be regarded as a disadvantage.
In this study, we have presented two approaches to the simulation
The data used for our study were a subset of the Jura data set
of a multivariate data set which rely on successful spatial
and the variables chosen are linearly correlated, an aspect which
decorrelation of the raw data, standardised to unit variance and is important as it is known that MAF does not cope well with
mean zero, prior to simulation. The approach differs from the non-linear relationships between the data (see for example
standard approach in that the data were not transformed to normal Rondon and Tran, 2008). The four variables in our subset
scores prior to the application of the decorrelation algorithms. provided a mix of isotropic and anisotropic data, making the use
However, in contrast to the implementation in Bandarian (2008), of an algorithm such as sequential Gaussian co-simulation
rather than using a direct sequential simulation algorithm, the unattractive as a parsimonious model of coregionalisation would
factors were simulated using sequential Gaussian simulation. have had to be chosen for the simulation. For both methods the
Thus, it was necessary to the convert each factor to normal scores anisotropy in the data is broadly reproduced, both at the
and back-transform after completion of the simulations, the latter univariate and the bivariate level, even though omnidirectional
operation being automatic in sequential Gaussian simulation. The semi-variogram matrices were used for the derivation of the
choice of a standard algorithm was largely prompted by its ready factor transformations.
E M BANDARIAN et al
FIG 11 - Directional experimental semi-variograms for standardised cobalt compared to mean directional semi-variograms for Alternating
Columns Diagonal Centres simulations and minimum/maximum autocorrelation factors simulations.
FIG 12 - Directional experimental semi-variograms for standardised nickel compared to mean directional semi-variograms for Alternating
Columns Diagonal Centres simulations and minimum/maximum autocorrelation factors simulations.
The results show that both decorrelation methods are viable Manton, J H, 2005. A centroid (Karcher mean) approach to the joint
for the simulation of the Jura data set. The results indicate that approximate diagonalisation problem: The real symmetric case,
ACDC performs at least as well as MAF for this set. Both Digital Signal Processing,16(5):465-478.
decorrelation methods considered use approximate diagonalisation Rondon, O and Tran, T T, 2008. Multivariate simulation using min/max
via a congruence transformation. The main difference lies in the autocorrelation factors: Practical aspect and case studies in the
mining industry, in Proceedings of the Eighth International
assumption that a 2SLMC describes the LMC for the data in the Geostatistics Congress (eds: J Ortiz and X Emery), pp 269-278.
case of MAF, which is not needed for ACDC. The ACDC
Soares, A, 2001. Direct sequential simulation and cosimulation,
method thus gives the user greater flexibility and is as easy to use Mathematical Geology, 31(2):155-173.
as the MAF decorrelation method. Given the performance for the Tercan, A E, 1999. Importance of orthogonalization algorithm in
Jura data, the ACDC method for decorrelation has the potential modelling conditional distributions by orthogonal transformed
to be used successfully with data sets whose LMC is not fully indicator methods, Mathematical Geology, 31(2):155-174.
characterised by two structures. Tran, T T, Murphy, M and Glacken, I, 2006. Semivariogram structures
used in multivariate conditional simulation via minimum/maximum
REFERENCES autocorrelation factors, in Proceedings XI International Congress,
IAMG, Liege.
Bandarian, E M, 2008. Linear transformation methods for multivariate
Vargas-Guzman, J A and Dimitrakopoulos, R, 2003. Computational
geostatistical simulation, PhD thesis (unpublished), Edith Cowan
properties of min/max autocorrelation factors, Computers and
University, Perth.
Geosciences, 29:715-723.
Bandarian, E M, Bloom, L M and Mueller, U A, 2008. Direct
Wackernagel, H, 2003. Multivariate Geostatistics, third revised edition,
minimum/maximum autocorrelation factors for multivariate
387p (Springer-Verlag: Berlin).
simulation, Computers and Geosciences, 34:190-200.
Yeredor, A, 2000. Approximate joint diagonalization using
Bandarian, E M and Mueller, U A, 2008. Reformulation of MAF as a
non-orthogonal matrices, in Proceedings International Workshop on
generalised eigenvalue problem, in Proceeding of the Eighth
Independent Component Analysis and Blind Source Separation
International Geostatistics Congress (eds: J Ortiz and X Emery),
(ICA2000), pp 33-38.
pp 1173-1178.
Yeredor, A, 2002. Non orthogonal joint diagonalization in the least square
sense with application in blind source separation, IEEE Signal
simulation framework and application in a multivariable deposit,
Processing, 50(7):645-648.
(ed: R Dimitrakopoulos), pp 345-354 (The Australasian Institute of Yeredor, A, 2004. ACDC: Approximate joint diagonalisation (in the
Mining and Metallurgy: Melbourne). direct Least-Squares sense) of a set of Hermitian matrices, using the
iterative ACDC algorithm [MATLAB code]. Available from:
Datta, B N, 1995. Numerical Linear Algebra and Applications, (Brookes/
<www.eng.tau.ac.il/~arie/>. [Accessed: 3 May 2009].
Cole Publishing Company: Pacific Grove).
Xie, T, Myers, D E and Long, A E, 1995. Fitting matrix-valued variogram
Desbarats, J A and Dimitrakopoulos, R, 2000. Geostatistical simulation
models by simultaneous diagonalization (Part II: Application),
of regionalized pore-size distributions using min/max autocorrelation
Mathematical Geology, 27:877-888.
factors, Mathematical Geology, 32(8):919-942.
483 p (Oxford University Press: New York).
HOME
Conditional Wavelet Based Simulation of Non-Stationary

Geologic Fields Using Geophysical and Model Analogue
Information
E Gloaguen1 and R Dimitrakopoulos2
ABSTRACT new approach allows the combination of the easy-conditioning of

pixel-based algorithms with the ability to reproduce shapes of
Non-stationarity and scale dependencies are critical topics when modelling
spatial phenomena of complex geological patterns. A two-dimensional
object-based techniques, without relying on excessive cpu
conditional simulation based on wavelet decomposition of geophysical demand (Strebelle, 2002). Recently, research on developing
and model analogue data is proposed for simulating non-stationary simulation methods with multiple point statistics uses
geological patterns. The proposed method requires that geophysical data convolution of the training image with spatial filters yielding to
are available on the entire area under study, and that we have an idea of filter scores that permits the clustering of similar spatial patterns
the small scale patterns of the geology (model analogue). Then, the in the filter score space (Zhang, Switzer and Journel, 2006).
method utilises the wavelet transform of the geophysical data and of the The local multiple point statistics is captured from the training
model analogue to decompose both of them into wavelet coefficients at image, and reproduced by simulating patterns. Hence, the complex
different scales, and then quantifies their spatial dependencies. Using multiple-scale geological structures can be reconstructed in the
cokriging, the approximation coefficients of the variable under study are
simulation grid, conditional to a variety of subsurface data such
inferred from the approximation coefficients of the geophysical data.
Joint simulation of the wavelet coefficients of the variable under study is as well data and seismic survey. A major limitation of existing
employed together with available soft conditioning data. The conditionally high-order statistics simulations is that they do not consider or
cosimulated wavelet coefficients are back transformed generating carry consistent relations between patterns at different scales, the
realisation of the attribute under study. The proposed method is applied topic of ongoing research.
on a synthetic channel example. Two models are generated, one represent- The relation between patterns at different scales can be assessed
ing the ‘real geological model’ and the other one, is the model analogue. through discrete wavelet transform or DWT (Portilla and
The model analogue represents the same stratigraphy (channels, number of Simoncelli, 2000). The properties of DWT coefficients fully
geological units) but differs by the spatial location and by the proportion of characterise an image in terms of scale dependencies, features and
the channels. The synthetic geophysical data consists in full coverage of
statistics (Mallat, 1989). DWT decomposes an image into wavelet
the model under study. The model analogue is transformed into a physical
property model in order to compute the forward modelling of the
coefficients that are regrouped in directional sub-bands related to a
geophysical property. The results show that wavelet simulation of the given scale (or level) of an image and allows both multiple scale
geological field shows good agreement with the synthetic one both in analysis and geostatistical modelling (Flandrin, 1992). The
terms of amplitudes and shape of the channels. dependencies of wavelet coefficients at different scales, namely,
high-order dependencies of wavelet coefficients, are studied for
INTRODUCTION statistical image modelling with applications to texture analysis
and texture synthesis (Crouse, Nowak and Baraniuk, 1998). These
The modelling and assessment of spatial uncertainty of the techniques use hidden Markov models (HMM) and Gaussian
attributes of geological phenomena are frequently based on the mixture model (GMM). Gaussian mixture model maximised the
stochastic simulation of random fields conditional to available model likelihood between the data and the mixture model to
data. Conventional simulation methods (Chilés and Delfiner, capture the marginal statistics. HMM attempts to take into account
1999; Goovaerts, 1997) represent the geological complexity in the interscale correlation that can be seen as a multidimensional
terms of second order statistics at a single scale, thus limiting GMM that applies to tree-structured Markov chains across scales
their ability to model more complex features (Guardiano and (Choi and Baraniuk, 2001). The limitations of these techniques
Srivastava, 1993). Here scale refers to the support size on which include their inability to allow direct conditioning to data, only the
a given pattern is defined. During the past years, substantial correlation between scales is taken into account and the likelihood
efforts have been made to develop new techniques that may take modelisation is an computer intensive iterative process. Using a
into account the so-called high-order spatial statistics (Mustapha different wavelet based approach Tran, Mueller and Bloom (2002)
and Dimitrakopoulos, 2010, in this volume) including multiple- have shown that unconditional Haar wavelet simulations can give
point statistics (Guardiano and Srivastava, 1993; Strebelle, results equivalent to the classical sequential Gaussian simulations
2002), computer graphic approach (Zhang, Switzer and Journel in the case of multi-Gaussian data.
2006; Arpat and Caers, 2004), Markov random field sequential This paper proposes a method for direct conditional simulation
simulation (Daly, 2004) and directional metropolis for non-linear in two dimensions, based on the DWT coefficient of a indirect
likelihood posterior updating (Tjelmeland and Eidsvik, 2004). As geophysical measurement and geological analogue. First, the
the two last methods are based on data, the multiple point DWT of the geological and corresponding geophysical analogues
geostatistics approach (Guardiano and Srivastava, 1993) uses a is used to capture the high-order dependencies between features at
training image instead of a two-point variogram to account for different scales between the geological and geophysical data. Also,
geological information. The training image describes the the high-order dependencies of the geological features is explored
geometrical facies patterns believed to be present in the through covariances and cross-covariances modelling of the DWT
subsurface and reflect a prior geological/structural concept. This coefficients. The covariance captures the spatial relation between
directional wavelet sub-bands representing a given resolution and
1. Institut National de la Recherche Scientifique, 490 rue de la the cross-covariances capture the frequency relation between the
Couronne, Québec QC G1K 9A9, Canada. directional wavelet sub-bands at different scales. Subsequently,
Email: [email protected] realisation of the DWT coefficients are joint simulated using a
2. FAusIMM, Professor and Director, COSMO – Stochastic Mine sequential Gaussian cosimulation approach (Verly, 1994). Finally,
Planning Laboratory, Department of Mining and Materials the DWT coefficients are backtransformed in the space of the
Engineering, McGill University, Montréal QC H3A 2A7, Canada. attribute of interest using the inverse DWT. A key aspect of the
Email: [email protected] method is that it allows conditioning to any measured data (hard
E GLOAGUEN and R DIMITRAKOPOULOS
conditioning) or conditioning only at a given scale based on

indirect information (soft conditioning).
TWO-DIMENSIONAL DISCRETE WAVELET

TRANSFORM
In this section the two-dimensional DWT is presented and
important scaling properties of wavelet coefficients is revisited in
the context of geostatistical simulation. Then, conditioning to
hard and soft data from analogue petro-physical scenario is
presented.
Two-dimensional wavelet transform revisited

By construction, wavelet transform provides both spatial and
frequency-domain information about a regularly sampled grid
such as an image (Mallat, 1989). Wavelet analysis procedures
involve the decomposition of an image in terms of a set of
translated and dilated wavelet functions ΨB (where B = {H, V, D})
and scaling functions (φLL) that form an orthonormal basis of the
Hilbert space L2(2) of square integrable functions (Fan and Xia,
2003).
Similarly to the two-dimensional Fourier Transform, the
columns of an image are analysed with the scaling functions
FIG 1 - Three discrete wavelet transform scales. All the subgrids
(φLL) and bandpass filters ΨB, and the rows of each of these
represent the same area. Arrows point from the parent (scale j-1)
resultant images are analysed again with each of (φLL) and ΨB.
toward the child (scale j). Hj, Vj and Dj are the directional wavelet
This filtering step results in four sub-bands for each scale of
detail coefficients at scale j and a3 is the approximation at scale
decomposition j, where j=1,...,J; the approximation characterising J=3.
the omnidirectional low frequency variations and three details
representing the local directional variations in three directions,
horizontal, vertical and diagonal (H, V, D). tree structure (Kumar and Foufoula-Georgiou, 1997). This
property is important as DWT decompose the image in sets of
For a given scale decomposition j=1,..,J, DWT of an image
oriented coefficient at different scale allowing to characterise
x(s,t) of size N N, where s and t are the Cartesian coordinate
possible geometric anisotropies at different scales. Note that each
axis, is given by (Fan and Xia, 2003; Crouse, Nowak and
sub-band represents the same area as the original training image
Baraniuk, 1998):
x(s,t).
Nj −1 J Nj −1 Property 2: DWT allows taking into account the opposite
x(s,t)= (s,t)+ Z Bj , k , i ψBj , k , i ( s , t ) (1) relation between resolution and support size. As a reminder, this
k, i = 0 B = {H ,V , D} j = 1 k , i = 0 principle tells that the more precise the reslolution is, the smaller
the support size should be, and vice versa. For example, the
where: small grids on the diagonal coefficients (Figure 1) represent the
same area but the resolution of each child is four times less than
−j the resolution of its parent. The explicit modelling of the
φ LL
J , k, i = 2
2 φ( 2 − j s − k, 2 − j t − i ) and cross-covariances between features at different scales is
−j particularly interesting for down-scaling purposes.
N
ψ BJ , k , i = 2 2
ψ B ( 2 − j s − k, 2 − j t − i ) with Nj = . Property 3: In theory, the DWT should decompose an image
2j with uncorrelated wavelet coefficients representing the most
important features of an image, but Crouse, Nowak and Baraniuk
a J, k, i = x ( s, t )φ j 0 , k , i dsdt are the approximation coefficients and (1998) demonstrated that there are still considerable dependencies
between scale, as observed from the characteristics of wavelet
Z Bj , k , i = x ( s, t )ψ Bj , k , i dsdt are the (k,i)th wavelet coefficients of
coefficient distribution (Fan and Xia, 2003). This proves that there
sub-band B in scale j. is still scale dependency in the wavelet coefficient. The
dependencies between scales are classified in two groups:
In the following paragraphs, the important properties of wavelet
intrascale and interscale as discussed in the next sections.
decomposition are defined in the context of two-dimensional
simulation with scale dependency preservation. To make the By definition (Equation 1), at a given scale, the three directional
notation lighter, the wavelet coefficients or sub-band coefficients decompositions are collocated and represent the directional
at a given level j, Z Bj , k , i are noted Hj, Vj and Dj. Each level component of an image at a given frequency (for example H3, V3,
represents a given scale of the image. Figure 1 shows the two- D3 and a3 in Figure 1). The cross-variance between the sub-bands
dimensional decomposition tree of an image x(s,t) for J = 3 belonging to a same scale characterises the spatial link between
the directional features of an image at this scale. Equation 1 also
decomposition levels. The link between the different scales is
represents the link between the smallest structures of an image and
represented by the arrows, pointing toward the scale j from the
the largest structures being consistent with the trade off relation
generating scale j-1. As an example, in Figure 1, each decom- between resolution and support size. The cross-covariance
position scale produced one approximation coefficient map (a3) between two sub-bands belonging to two different scales
and three directional detail coefficient maps (H3, V3, D3). characterises the spectro-spatial information shared between the
two sub-bands (for example H3 and V1 in Figure 1). The wavelet
Practical aspects of the wavelet coefficients coefficients at different scales are nested. Consequently, starting
from the coarsest level to a point in the finest level is similar to
Property 1: DWT provides with a set of directional and zooming without conflict between scales. This property is
frequency dependent coefficients that can be represented on a particularly important for scaling purpose (Mallat, 1989).
CONDITIONAL WAVELET BASED SIMULATION OF NON-STATIONARY GEOLOGIC FIELDS
Intra- and interscale cross-variances of this particular spatial each sub-band, the cross-variograms between sub-bands belonging
decomposition is a direct link to high-order statistics (Portilla to both same scale or not and any known wavelet coefficient data.
and Simoncelli, 2000) as each point at the coarsest scale is In the following paragraph, the sequential Gaussian cosimulation
related to each of its generating points belonging to the finer is presented in the two-dimensional wavelet coefficient frame-
scales. Once the cross-covariances of the inter- and intrascales work. The detailed theory of Sequential Gaussian CoSimulation
are modelled, the spatial statistics are then of the order of the (SGCS) can be found in Verly (1994).
number of level times three, the number of sub-bands.
For mathematical simplicity, we present only the estimation of
Inverse discrete wavelet transform (IDWT) the wavelet coefficients belonging to a same scale J=1 at location
x0. There are four covariables to estimate, Za1,ZH1, ZV1 and ZD1.
Inverse discrete wavelet transforms are used to back transform
conditionally cosimulated wavelet coefficients to realisations of Assuming the variables are joint multinormally distributed and
the attribute under study. IDWT is the back transformation from considering the already measured or previously simulated matrix
the wavelet space to the original space. Inverse discrete wavelet Z = [ Znd
a 1 H1 D1 V 1 T
Znd Znd Znd ] of size nd known values by four variables,
transforms can be performed by inverting the orthogonal matrix
that transforms the original coefficients into the coefficients in the conditional distribution of Zxa01 ZxH01 ZxD01 ZVx 01 at x0 is defined by
−1
the wavelet basis. In practice, the IDWT is computed starting its conditional mean m( Zxa01nd , H 1, D 1
) = C x 0 nd C ndnd Z and its
from the coarsest scale and combining the approximation and the a 1 , H 1 , D 1 ,V 1
conditional variance C ( Zx 0 nd −1
) = C x 0 x 0 − C x 0 nd C ndnd C ndx 0 .
details into the approximation at the next finer scale, and so on.
Where DWT involves filtering and downsampling, the IDWT where:
consists of upsampling and filtering. Downsampling by p is
defined as taking every pth sample, starting with sample 1. Up Cx0x0 is the matrix of covariance of the coefficients
sampling is the process of lengthening a signal component by
inserting zeros between samples. Cx0nd = Cndx0T is the covariance matrix between coefficients to
estimate and the already measured coefficients
Training image – A model of scale dependency Cndnd is the covariance matrix between already
The idea of modelling the scale dependency of an attribute from measured coefficients
a training image was first introduced in geostatistics by A realisation of the conditional random vector is given by
Guardiano and Srivastava (1992). The training image describes (Rubinstein, 1981):
the geometrical facies patterns believed to be present in the
subsurface and reflect a prior geological/structural concept. −1
Zxa01nds
H 1 , D 1 ,V 1
≈ C xondsC ndsnds Z + Lx 0 nds e (2)
The needs for a training image in the proposed DWT
simulation comes from three sources:
where:
1. The scale dependency may not be present in the finite set of
measured points. e is a vector of four independent standard Gaussian
2. Discrete wavelet transform works only on full and regular values
grids.
Lx0|nd is the lower triangular matrix of the Cholesky
3. In many cases, the specialist has a clear understanding of decomposition of C xa01ndH 1 , D 1 ,V 1
the global features present in the attribute under interest.
For example, it can be a geological model inferred from If only a subset nds of the nd conditioning data is considered
indirect geophysical measurements or from geological based on the neighbourhood of x0, then applying screen effect
knowledge. approximation, Equation 2 becomes:
To summarise, DWT allows computing the directional
spectro-spatial characteristics of a training image. DWT and Zxa01nds
H 1 , D 1 ,V 1 −1
≈ C x 0 ndsC ndsnds Z + Lx 0 nds e (3)
IDWT is a bijonctive process without loss. Conditional simulation
of the wavelet coefficients will allow generating set of wavelet
coefficients resulting in an image that should have the same where Zxa01nds
H 1 , D 1 ,V 1
is a 4 × 1 vector of containing the simulated
wavelet domain characteristics as the original training image.
values of Zxa01 , ZxH01 , ZxD01 and ZVx 01 at x0. They all respect the auto-
In the next section, we describe how to cosimulate the wavelet
coefficients in the wavelet space and back transform to the and cross-covariances so as the conditioning data.
original space to produce conditional simulations of the attribute
of interest preserving its scale dependencies. Intra- and interscale covariance
CONDITIONAL SIMULATIONS WITH DISCRETE To enhance the practical application of the method, the full
WAVELET TRANSFORM IN TWO DIMENSIONS modelling of DWT coefficient auto- and cross-covariances is
replaced here with the hypothesis of intrinsic correlation. Then
In this section, a conditional simulation algorithm of wavelet the cross-covariance is equal to simple correlation (Chilés and
coefficients is presented. First, Sequential Gaussian CoSimulation Delfiner, 1999). Two cases need be considered, intra- and
(SGCS) (Verly, 1994) is briefly outlined and then, the algorithm is
interscales, as the scales have different supports. The intrascale
explained in detailed for wavelet coefficient simulation.
coefficients are collocated and are on the same support, hence,
Sequential Gaussian cosimulations of discrete the simple correlation r is straightforward. For example, the
correlation between ZH1 and ZV1 with zero mean is given by:
wavelet transform coefficients
In the proposed method, the approximation and the detail coeffi- cov( Z H 1 , ZV 1 )
cients are treated as regionalised covariables. The cosimulation r = (4)
Z H 1 , ZV 1
algorithm must reproduce the distribution and the variogram of var( Z H 1 )var( ZV 1 )
where cov and var stand for covariance and variance, respectively. CASE STUDY
In the case of interscale, the variables belong to two different
scales having different spatial sampling. Similarly to the intrascale In order to explore our approach, the method presented above is
model, the cross-covariance between the inter-scale coefficients is applied to a synthetic case study where both geophysical and
computed using the simple correlation but spatial sampling has to geological data are known. Geological data (primary one)
be taken into account. The spatial relation between the wavelet corresponds to a channelled porosity field and the geophysical
coefficients belonging to different scales is represented by boxes data (secondary soft data) corresponds to the acoustic impedance
in the diagonal coefficient scales in Figure 1. Each box represents field (AI) measured on the geological model. Note that AI is
the same area but, for example, the resolution of the box in D1 is computed by multiplying the seismic velocity times the density
four times the resolution of the box in D2. In order to compute the of the probed material. Obviously, in the real data set, only the
correlation using relation 3, the values inside the highest resolution geophysical data are available. Figure 3 shows the synthetic
grid are averaged at the under sampled grid resolution. geological model we want to retrieve and Figure 2 shows the
corresponding geophysical model. In order to be as close as
Conditioning possible to real data acquisition, we added a Gaussian noise on
the geophysical data shown in Figure 2. Figure 4 shows the
As previously presented, DWT is a linear decomposition of an histogram of the noise free and the noisy AI (Figure 2). The
image. This allows the conditioning to existing soft or hard data standard deviation of the noise free AI is 275 m/s, whereas it is
to be implemented within the above simulation algorithm. 410 m/s for the noisy data. As can be seen on the scatter plot
In the soft conditioning case, the conditioning data (usually between noisy AI and the porosity field in Figure 5 the relation
geophysical data that shows systematic coverage over the whole between AI and porosity is weakly non-linear especially for
studied area, but with a poor resolution), measured at large scale in porosity below 25 per cent.
a regular grid, influence only the regional trend of the variable to
be simulated. In the case of wavelet coefficients simulation, soft
conditioning is obtain from converting all the approximation
coefficients of the secondary variable as conditioning data for the
approximation coefficients of the primary. To convert the
secondary data wavelet coefficients into primary ones, we produce
a geological analogue that represents the different coarse scale
textures present in the studied ground. This geological model is
translated in terms of geophysical properties using known
petro-physical relationships or in situ calibrated relations typically
obtained from borehole logs and cores. Both geological and
geophysical analogue are subjected to wavelet transformation.
Geophysical measurements give information on the large-scale
features of the ground and following this idea, focus is placed only
on the coarsest level of the wavelet coefficients. Statistical
relationships between the coarsest wavelet coefficients of both
geological and geophysical analogues are then studied. As noted
in the case study that follows, one of the most interesting
observations made during data analysis is that the weak
non-linearity between both analogues in the original space
becomes highly linear at the coarsest wavelet space. This occurs
principally because of the physics of the phenomenon – that is the FIG 2 - Measured acoustic impedance.
geophysical data are coarse scale representation of the geological
model. Due to the strong linearity of the relation between coarse
wavelet coefficients of both geophysical and geological analogues
and its ease to incorporate hard data, cokriging is used to infer
‘true’ geological coarse wavelet coefficients based on measured
geophysical wavelet coarse coefficients.
Quantification of the quality of the simulation

In this section, a quantitative way to quantify the quality of the
simulation is discussed.
The Julesz’ conjecture (Julesz, 1962) states that there exists a
set of statistical measurements such that two images are
perceptually indistinguishable if and only if they are drawn
from the same random function matching those statistics. In the
meantime, Julesz constructed examples, by creating pairs of
textures with the same nth-order pixel statistics that were
strikingly different in appearance. Also, it would be unlikely,
that a pair of images that produce identical statistical estimates
looks very different. In the present paper, the main
characteristics of the wavelet based simulations are that they
have to respect the covariances, the cross- covariances between
wavelet sub-bands and any conditioning data. Consequently, FIG 3 - Real porosity model.
performance of the method will be tested by applying DWT on
the simulated images and by computing their covariances and
cross-covariances of the DWT coefficients. Simulated images Based on our knowledge of the geology, we build a geological
wavelet coefficient statistics should reproduce the same analogue (Figure 6). The geological analogue differs from the
statistics as the training image wavelet coefficients. real field (Figure 3) by both the orientation of the channels
FIG 4 - Noisy and noise free measured acoustic impedance. FIG 7 - Analogue acoustic impedance computed on porosity field.
As shown on Figure 1, third level wavelet decomposition is

applied on the measured geophysical data and on both geological
and geophysical analogues (Figure 8). The choice of the wavelet
decomposition scale is physically obtained when the correlation
between the geological and geophysical wavelet approximation
analogue is maximised. This approach allows the fusion of the
geological and geophysical data at the appropriate scale. Figure 9
shows the scatter plot between approximation coefficients of the
AI and porosity analogue. The relation between both data is now
very linear in the wavelet space and the correlation is as high as
0.981. As a comparison, the correlation between the measured AI
and its corresponding porosity is 0.656 whereas the correlation
between their wavelet approximation coefficients is 0.956. This
is due to the fact that geophysical data carries information on the
coarse scale of the probed ground. The approximation coefficients
of the porosity are cokriged using the modelled variogram of the
measured AI, the correlation coefficient between analogue AI and
porosity and the intrinsic hypothesis.
Finally, the variograms and cross-variagrams of each
FIG 5 - Scatter plot between acoustic impedance and porosity. sub-bands belonging to the same level or not are modelled. Using
the full linear model of corregionalisation and the porosity
approximation coefficients computed from the measured AI, the
porosity wavelet coefficients are cosimulated, starting from the
coarsest scale to the finest. When all the coefficients of all scales
are cosimulated, the wavelet coefficients are back-transformed in
the original space using IDWT. Figure 10 shows one simulated
porosity field. Compared to the real model shown in Figure 3, the
shape and intensity of the porosity of the background and within
channel is reasonably well reproduced except for some
numerical artifacts essentially due to the discretisation of the
analogues. Also, the correlation between the simulated porosity
field and the porosity model shown in Figure 3 is around 0.9. In
addition, even if the measured AI were really noisy, the
algorithm acts as a fine scale low-pass filter. It means that each
scale of an image can be filtered at its given scale. Finally, Figure
11 shows the e-type of 20 simulations. All the random numerical
artifacts due the discretisation vanished.
CONCLUSIONS
FIG 6 - Analogue porosity field. In this paper, it is demonstrated that DWT coefficients analysed
with well-known geostatistical methods can generate simulations
using modelling of the wavelet coefficient statistics of the field to
(Figures 6 and 7) and histogram (Figure 8). A geophysical AI simulate. It can be shown that the wavelet scale statistics are
field is computed from the analogue porosity field using known reproduced at least in terms of inter- and intrascale cross-
petro-physical relationships (Pride, 1994). The corresponding covariances. This proves that some scale dependencies are
analogue AI field is shown in Figure 7. reproduced in the simulated images. Realisations generated using
FIG 9 - Scatter plot between the analogue acoustic impedance

approximation coefficients of Figure 7 and the analogue porosity
approximation coefficients of Figure 6.
FIG 10 - One simulated porosity field.
FIG 8 - Three-level wavelet decomposition of: (a) measured

acoustic impedance (Figure 2), (b) porosity analogue (Figure 6)
and (c) acoustic impedance analogue (Figure 7). The
approximation coefficients are in the upper left box.
the proposed method reproduce the non-stationarity observed on

the geophysical data, the conditioning data, the wavelet
coefficients and their spatial dependencies. The coarse to scale
simulation approach allows simulating using both hard and soft
data and both are fitted at their appropriate scale. The methods still
work even if the analogue shows different spatial anisotropy in the
analogue. This property prevents the use of non-friendly rotation
matrices. The methodology can be easily extended to more than
one covariate geophysical data. The only change from the
proposed example is the cokriging step of the approximation
coefficient that becomes an n-variable cokriging instead of two FIG 11 - E-type of 20 realisations.
variables. A limitation of the method is that the training image has Guardiano, F and Srivastava, R M, 1993. Multivariate geostatistics:
to contain the same wavelet statistics as the field to be simulated. beyond bivariate moments, in Geostatistics Troia (ed: A Soares),
Hence, the difficulty is to find appropriate training images. In 92:133-144.
other hand, in the soft conditioing case, the method doesn’t require Julesz, B, 1962. Visual pattern discrimination, IRE Trans Info Theory, IT,
the training image to have the same trend direction. volume 8, pp 84-92.
Kumar, P and Foufoula-Georgiou, E, 1997. Wavelet analysis for
geophysical applications, Reviews of Geophysics, 35(4):385-412.
ACKNOWLEDGEMENTS Mallat, S, 1989. Multifrequency channel decompositions of images and
Thanks are in order to the industry members of the COSMO wavelet models, IEEE Transaction on Acoustic Speech and Signal
Processing, 37:2091-2110.
Laboratory: AngloGold Ashanti, Barrick, BHP Billiton, De
Mustapha, H and Dimitrakopoulos, R, 2010. Geologically enhanced
Beers, Newmont, Vale and Vale Inco, as well as NSERC
simulation of complex mineral deposits through high-order spatial
(Discovery Grant 239019), the Canada Research Chairs Program cumulants, in Advances in Orebody Modelling and Strategic Mine
and CFI are gratefully acknowledged. Planning I (ed: R Dimitrakopoulos), pp 309-320 (The Australasian
REFERENCES Portilla, J and Simoncelli, E P, 2000. A parametric texture model based
on joint statistics of complex wavelet coefficients, Int J of Comput
Arpat, G and Caers, J, 2004. A multiple-scale, pattern-based approach to Vision, 40(1).
sequential simulation, in Geostatistics Banff 2004, Springer, Pride, S, 1994. Governing equations for the coupled electromagnetics and
Dodrecht (eds: O Leuangthong, and C V Deutsch), pp 255-264. acoustics porous media, Physical Review B, 50:15678-15696.
Chilès, J-P and Delfiner, P, 1999. Geostatistics, Modelling Spatial Rubinstein, R Y, 1981. Simulation and the Monte-Carlo method, 272 p
Uncertainty, 385 p (Wiley: New York). (Wiley: New York).
Choi, H and Baraniuk, R, 2001. Multiscale image segmentation using Strebelle, S, 2002. Conditional simulation of complex geological structures
wavelt domain hidden Markov models, IEEE Transactions on Image using multiple-point statistics, Mathematical Geology, 34:1-21.
Processing, vol 10, pp 1309-1321.
Tjelmeland, H and Eidsvik, J, 2004. Directional Metropolis: Hastings
Crouse, M S, Nowak, R D and Baraniuk, R G, 1998. Wavelet-based updates for posteriors with non linear likelyhood in, Geostatistics
statistical signal processing using hidden Markov models, IEEE Banff 2004, pp 95-104 (eds: O Leuangthong and C V Deutsch)
Transactions on Signal Processing, 46(4):886-902. (Springer: Dodrecht).
Daly, C, 2004. Higher order models using entropy, Markov random fields Tran, T, Mueller, U A and Bloom, L M, 2002. Multi-scale conditional
and sequential simulation, in Geostatistics Banff 2004 (eds: O simulation of two-dimensional random processes using Haar
Leuangthong, and C V Deutsch), pp 215-224 (Springer: Dodrecht). wavelets, Proceedings of the GAA Symposium, pp 56-78.
Fan, G and Xia, X-G, 2003. Wavelet-based texture analysis and synthesis Verly, G W, 1994. Sequential Gaussian cosimulation: A simulation
using hidden Markov models, IEEE Trans on Circuits and Systems – method integrating several types of information in, Geostatistics
Fundamentals, Theory and Applications, 50(1):106-120. Troia, 1992 (ed: A Soares), pp 85-94 (Kluwer Academic Publishers:
Flandrin, P, 1992. Wavelet analysis and synthesis of fractional Brownian Dordrecht).
motion, IEEE Trans on Information Theory, 35:197-199. Zhang, T, Switzer, P and Journel, A, 2006. Filter-based classification of
Gloaguen, E and Dimitrakopoulos, R, 2009. Two-dimensional conditional training image patterns for spatial simulation, Mathematical Geology,
simulations based on the wavelet decomposition of training images, volume 38, no 1, pp 63-80.
Mathematical Geosciences, 41(6)679-701.
Goovaerts, P, 1997. Geostatistics for natural resources evaluation, 483 p
(Oxford University Press).
HOME
Optimal Life-of-Mine Scheduling for a Bauxite Mine

M Zuckerberg1, J van der Riet2, W Malajczuk3 and P Stone4
ABSTRACT combination of day/night and summer/winter. Also, over the

This paper describes a new life-of-mine planning optimisation software
mine life, a number of critical decisions need to be made
tool named Bodor (Boddington Optimiser) which has been developed for regarding the timing of investment in new crushing and overland
BHP Billiton’s Boddington Bauxite mine in South-Western Australia. conveyor infrastructure, or the merits of recycling existing
Bodor minimises pretax net present cost (capital and operational) over a crushers. Traditionally, mine planners have used some sensible
specified mine life, and is applied to a mine model consisting of bauxite heuristics to guide a scheduling model implemented in a popular
pods predesigned to a fixed cut-off grade, directly feeding a refinery with commercially available mine planning and scheduling tool.
bauxite that meets grade and throughput targets in each period over the These heuristics include delaying new crusher and conveyor
life-of-mine. Bodor has been very successful in its application to the infrastructure capital expenditure as long as possible and
Boddington mining operation, producing a new life-of-asset mine plan
which delivers a significant reduction in net present cost. This value add minimising operating haulage costs in obvious ways.
is achieved primarily through a better timing of orebody exploitation Unfortunately, these heuristics are not always reliable due to the
which optimally trades-off lower haulage costs against some capital costs natural trade-off between capital and operating costs, viz: by
that are brought forward. A core aspect of Bodor’s utility lies in delaying the expenditure of capital associated with the opening
managing various complex environmental constraints in determining the of new envelopes, the haulage costs will unavoidably increase as
optimal extraction schedule. These constraints lead to minimising noise the cheaply recoverable pods close to the existing crusher are
disruption to local landholders through accounting for seasonally progressively exhausted. Moreover, the confounding issues of
prevailing winds and day/night access controls, maximising the resource
meeting blend targets at the refinery gate and respecting
life through determination of the best blend strategies, and optimising the
mining truck fleet usage by balancing day and night deployment. environmental constraints add further complications. These
environmental constraints relate to the seasonality of prevailing
winds, the close proximity of landholders to some of the pods
INTRODUCTION (see Figure 3), and the desirability of opening, extracting, closing
The Boddington mine lies just south-east of Perth in and rehabilitating/remediating pods within a reasonable time
South-Western Australia and is a relatively low grade – average frame. Examples of the rehabilitation process are shown in
alumina grade of 31 per cent – but long-life bauxite resource, Figure 4.
which supplies the Worsley alumina refinery near Bunbury. It is a
joint-venture operation, 86 per cent owned by BHP Billiton,
ten per cent by Japan Alumina Associates and four per cent by
Sojitz Alumina. The bauxite ‘pods’ that constitute the resource of
some 700 million dry tonnes are spread over a lease area of
2700 square kilometres, and are clustered into several distinct
mining envelopes. Currently, only one of these envelopes is mined
and sent to the refinery via a 51 km overland conveyor. The Envelope 5
Boddington life-of-mine plan calls for the eventual development,
opening, extraction and remediation of pods in all mining
envelopes. Each envelope will be serviced by a crusher, with the
crushed ore being transported to the refinery via a network of
overland conveyors. Overburden and waste is stockpiled locally
and then used to remediate the shell of the exhausted pod. This
configuration, which is explained in more detail in the next section Envelope 4
(Bodor Asset Model), is illustrated in Figure 1 – the mining and
conveying operation is shown in Figure 2. Envelope 3
Mine planners at Boddington face a very complex task in

calculating the best extraction sequence for the resource over the Overland Conveyor
life-of-mine. Not only do refinery feed blend targets have to be
Crusher
met in each period, but the management of the mining truck fleet
is cost critical, given that many pods in the deposit can only be Envelope 2 Bauxite Pod
mined in the daytime, some in summer only and some in any
1. Principal Scientist, BHP Billiton Resource and Business

Optimisation, 180 Lonsdale Street, Melbourne Vic 3000.
Email: [email protected] Refinery
2. (Deceased) Formerly with: BHP Billiton Bauxite Alumina Technology

Centre, Gastaldo Road, Worsley via Collie WA 6225.
3. MAusIMM, Superintendent Mine Planning, Worsley Alumina Pty Envelope 1
Ltd, Boddington Bauxite Mine, Gastaldo Road, Worsley via Collie
WA 6225. Email: [email protected]
4. Manager Optimisation R&D, BHP Billiton Resource and Business
Optimisation, 180 Lonsdale Street, Melbourne Vic 3000.
Email: [email protected] FIG 1 - Bauxite mine asset model topology.
M ZUCKERBERG et al
Bodor (Boddington Optimiser) has been developed in-house at

BHP Billiton’s Melbourne Technology Centre to enable
Boddington’s mine planners to calculate a resource extraction
sequence over the life-of-mine that minimises net present cost
(capital and operational), meet all bauxite blend targets at the
refinery front gate, respect environmental and operational
constraints, and allow smooth utilisation of trucking resources. It
is a software product with a graphical user interface and a
mixed-integer-linear-programming optimisation engine, and is
used directly by Boddington mine planners to evaluate
life-of-mine plans rapidly – a single optimal life-of-mine plan
can be calculated overnight. In this paper, we will outline the
capability of Bodor and show how it has been used by
Boddington mine planners to develop optimised life-of-mine
plans. In the last section, we discuss several enhancements to
Bodor already in the development pipeline.
BODOR ASSET MODEL

The centrepiece of the Bodor asset model is the ‘pod’, which is a
distinct body of ore lying close to the surface and of modest size
– usually less than two million tonnes. Bodor is specifically
designed for situations in which there are numerous pods that are
relatively small in size – of the order of several hundred – so that
no single pod would require more than approximately three years
to mine. Bodor consequently focuses on the sequence in which
pods are to be excavated, rather than on the block extraction
sequences of the individual pods. At Boddington, some of the
pods are composed of a collection of ‘subpods’, effectively
simple mining phases. Subpods must be developed in a
designated sequence as illustrated in Figure 5. Each subpod is
assumed to be homogeneous in all respects and hence can be
extracted in the optimised schedule in any proportion. Subpods
FIG 2 - Photos of the Boddington mining operation, showing have a physical mining precedence, one to another. The
truck/shovel excavation of shallow surface pods and the overland precedence rules are that subpods are numbered in their strict
conveyor to the Worsley refinery. order of possible extraction; viz, subpod 2 can only be extracted
after subpod 1 and so on. Figure 1 illustrates that each pod is a
member of a mining envelope. All pods in a given envelope
deliver their ore to a particular associated crusher. There is a
one-off capital cost associated with ‘opening’ an envelope. There
are also one-off capital costs associated with building the crusher
(or relocating it from another envelope) and the necessary
conveying system used at the envelope. No mining can occur
within an envelope before the envelope is opened. The cost to
open an envelope will depend on which and how many envelopes
have been previously opened and closed, as different envelopes
will share some infrastructure.
With regard to conveyers, at Boddington all ore mined from an
envelope undergoes primary crushing at the envelope’s local
crusher and is then conveyed to a secondary crushing point and
eventually on to the refinery. The conveying system need not
connect each envelope directly to the refinery. Rather, it is a
directed tree network of arcs (conveyer increments) with a root
node at the refinery and nodes representing the envelopes, such
that the ore of each envelope is conveyed via the arcs (conveyer
increments) of the tree on its way to the refinery. Specifically,
each envelope has a single downstream path leading to the
refinery and thus this entire path of arcs (conveyer increments)
must be constructed before the envelope may open.
BODOR OPTIMISATION PROBLEM
Objective
The Bodor objective is to calculate a feasible subpod extraction
schedule and set of exploitation decisions which will minimise
the net present cost and deliver the target tonnage and blend
FIG 3 - Position of noise-sensitive residences in relation to current quality to the refinery gate in each period whilst absolutely
mining operations at Boddington. obeying all constraints over a predetermined mine life. It was
OPTIMAL LIFE-OF-MINE SCHEDULING FOR A BAUXITE MINE
FIG 4 - Examples of land and forest rehabilitation at Boddington.
chosen to minimise net present cost, rather than maximise net • costs that are fixed and borne at the start of pod excavation,
present value, because a representative and fair revenue model
• costs that are fixed and borne in the development period
for the bauxite at the refinery gate was not available. before the start of pod excavation,
Exploitation decisions • costs that are fixed and borne at the conclusion of pod
excavation, and
Bodor is designed to answer such strategic life-of-mine planning • costs that are charged in proportion to the amount of the pod
questions as: extracted (able to be represented as $/tonne extracted).
• the determination of mining envelope’s opening and closing The planner must also provide Bodor with the asset model
times, topology, as well as all relevant constraints and targets applying
• whether crushers should be constructed or redeployed from over the prescribed mine life.
one mining envelope to another mining envelope,
Constraints
• whether or not the conveyer infrastructure for one of the
mining envelopes may be redeployed at a different mining The following constraints can be optionally applied by the
envelope, and planner in the determination of Bodor’s optimised schedule:
• the determination of optimal truck fleet sizes and optimally • Tonnage: target tonnage of alumina or bauxite delivered to
smooth allocations of trucks to the various classes of pods in the refinery gate per period.
each mining envelope. • Blend: upper and lower limits to the percentage of reactive
silica, total silica, extractable sulfate, extractable oxalate and
Inputs carbon in bauxite delivered to the refinery gate per period.
The pod model will contain all grade and cost information, • Maximum pod open time.
relating to the bauxite in-the-ground. In terms of costs, the pod • Subpods: for larger pods, subpods are mining phases that are
model should separately list the following: separately extractable and must be taken in a strict order.
Sub-Pod 4 Sub-Pod 3 Sub-Pod 2 Sub-Pod 1
FIG 5 - Bodor subpod topology.
M ZUCKERBERG et al
• Grade control drilled pods: those pods that have already been Although all parameters are configured via the graphical user
grade controlled can be given precedence in the schedule. interface, Bodor also provides a batch-processing facility –
planners may configure a number of optimisation scenarios and
• Earliest start and latest finish.
submit them to the batching utility for sequential processing
• Maximum mining rate. overnight.
• Maximum conveyor capacities.
• Noise and dust: these are day/night and summer/winter Two-phased approach
restrictions, which are applied to manage both noise and dust
Mine planners at Boddington currently calculate life-of-mine
pollution and to limit the general impact of mining
schedules on the basis of 18 single month periods, followed by
operations on adjacent landholders. Every hour available for
mining is classified either as ‘day’ or as ‘night’, and either as 14 quarterly periods, followed by between 25 and 35 yearly
‘summer’ or ‘winter’. Time can thus be divided into four periods. This is a very large number of periods for a mine
disjoint categories of summer day, summer night, winter day planning optimisation problem, and with no guidance as to
and winter night. Pods may be designated as available for when envelopes and sectors will open and close, this problem in
mining during only some of these four times. For example, a Bodor is not currently tractable. A 36 year schedule whose
pod may be ‘day only’ ‘summer only’, or ‘summer day and periods are 18 months each, however, is indeed tractable. This
winter night only’, etc. These constraints serve essentially to suggests that a two-phased approach can be used to render the
limit the amount of ore that may be extracted from such pods whole optimisation problem tractable. In the first phase, larger
in each period. period lengths with some constraint relaxations are used to
determine approximate opening and closing times for each pod
• Lead time: the user will specify the number of months and in the second phase the full optimisation problem is solved
required for the construction or relocation of conveyers and for the shorter period lengths by applying relaxed earliest start
crushers. No initially closed envelope will be allowed to open and latest finish constraints as informed by the first phase
before the requisite lead time elapses for constructing its results. Thus, the primary purpose of the first phase of
infrastructure. Furthermore, no envelope will be allowed to optimisation is to bind the opening and closing times of the
open with a relocated crusher or conveyer before the requisite envelopes and pods to within narrow, but not over-constraining,
lead time elapses after the closure of the envelope from which limits. We have tested Bodor extensively to confirm that the
the crusher or conveyer is being relocated. loss in value due to the implementation of this two-phase
• Smooth trucking: smooth and consistent utilisation of truck approach is minimal, in the order of fractions of one per cent of
resources can be enforced. the total net present cost.
Outputs
BODOR PERFORMANCE
Bodor’s optimiser engine will output an assignment of a fraction
between zero and one to each subpod and period, indicating the Bodor has been applied to the 30-year plan at the Boddington
fraction of the ore in the corresponding pod that is to be excavated Bauxite mine for the existing 3.5 Mt/a (alumina) operational
in the corresponding period. For example, Bodor may output the configuration. The Bodor schedule was compared to the existing
number 0.4 to subpod I period J, which will indicate that 40 per benchmark schedule, which was developed using manual guidance
cent of subpod I is to be excavated during period J, but Bodor will of the XPAC scheduling software. In the development of the
not identify which 40 per cent of the ore ought to be extracted. benchmark, the principle used to guide the XPAC scheduler was to
Bodor assumes the material in the subpod to be homogenous, so delay major capital expenditure as long as possible. In the end, the
that the 40 per cent of the subpod that was extracted will be benchmark schedule did not respect all constraints in all periods
assumed to possess the same grade parameters as the subpod as a absolutely. However, it represented the best schedule that mine
whole. Periods can be of any length, practically from one month to planners could construct with existing commercially available
several years as required by the planner. tools. In contrast, the Bodor schedule does respect all constraints
The optimiser will output a comprehensive schedule of all pod in all periods and furthermore delivered a net present cost that is
openings and extractions, as well as of the openings and closings five per cent lower than that of the benchmark. The majority of
of each of the envelopes and the construction dates of each of the this value is delivered by implementing a better mining strategy.
conveyer increments. For each envelope it will indicate whether Figure 6 shows that Bodor brings forward the opening of new
the envelope is to be configured with a new or a relocated envelopes and the associated capital charge, but balances this
crusher. Bodor will also determine the best truck fleet size for
through smaller truck fleets, reduced truck cycle times and more
each lease period, and output a comprehensive truck assignment
schedule that indicates how many trucks were assigned to each efficient utilisation of the truck fleet (Figure 7).
of the envelopes during each truck assignment interval.
CONCLUSIONS
BODOR CONFIGURATION
The success of Bodor, applied at the Boddington Bauxite Mine,
Bodor is designed around the CPLEX optimisation suite by demonstrates the value of state-of-the-art optimisation tools in
ILOG Inc (ILOG (2007)), and is deployed for use by operations the life-of-mine planning domain. Using a powerful MILP solver
mine planners on high-end laptop computers. It consists of three such as CPLEX combined with appropriately designed mining
major components: models and careful implementation of a two-phase optimisation
1. a user interface, where planners configure the mine asset strategy has allowed us to develop life-of-mine plans that respect
model and assign parameter values to the model and for the complex shorter-term operational constraints on a monthly basis
optimisation procedure, and also deliver a minimum net present cost bottom line.
2. the asset model, realised as a mixed-integer-linear-
programming (MILP) problem using CPLEX Concert REFERENCES
technology, and
ILOG Inc, 2007. ILOG CPLEX v10. Available from:
3. the standard CPLEX MILP optimisation engine. <http://ilog.com/products/cplex/>.
OPTIMAL LIFE-OF-MINE SCHEDULING FOR A BAUXITE MINE
Bodor
Exploitation Seq. by Envelope
ENVELOPE 1 TONNES ENVELOPE 2 TONNES ENVELOPE 3 TONNES ENVELOPE 4 TONNES ENVELOPE 5 TONNES ENVELOPE 6 TONNES ENVELOPE 7 TONNES
100%
80%
% Mining Envelope
60%
40%
20%
0%
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30
Period
Benchmark
Exploitation Seq. by Envelope
ENVELOPE 1 TONNES ENVELOPE 2 TONNES ENVELOPE 3 TONNES ENVELOPE 4 TONNES ENVELOPE 5 TONNES ENVELOPE 6 TONNES ENVELOPE 7 TONNES
100%
80%
% Mining Envelope
60%
40%
20%
0%
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30
Period
FIG 6 - Boddington envelope exploitation schedules.
Bodor
Noise Sensitive Tonnage Trend
SUMMER DAY TRUCK HOURS SUMMER NIGHT TRUCK HOURS WINTER DAY TRUCK HOURS WINTER NIGHT TRUCK HOURS
100%
% Total Utilised Truck Hours
80%
60%
40%
20%
0%
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30
Period
Benchmark
Noise Sensitive Tonnage Trend
SUMMER DAY TRUCK HOURS SUMMER NIGHT TRUCK HOURS WINTER DAY TRUCK HOURS WINTER NIGHT TRUCK HOURS
100%
90%
% Total Utilised Truck Hours
80%
70%
60%
50%
40%
30%
20%
10%
0%
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30
Period
FIG 7 - Boddington truck utilisation.
HOME
Coal Mining in the Hunter Valley – A Strategic Mine Planning

Case Study
C Wharton1
ABSTRACT basal seams and mining extents on tonnage and NPV was also
undertaken. The model was then later used to plan plant
Coal presents many engineering challenges and opportunities for
optimisation in a changing economic environment. The imminent impact expansions to cope with new coal fields, as well as decisions on
of carbon taxation and cleaner coal technology will add to these location, including truck haulage and/or conveyor options.
challenges. Ash content and other physical attributes dictate what types of
coal are available; however, the ability to wash and change the coal COAL MODELLING
characteristics, together with bypass and blending, leads to interesting
tonnage versus value opportunities – sophisticated tools are required to Inputs to the model
model and optimise these calculations. Long mine lives present a
challenge to net present value based optimisation – this paper will discuss The inputs to the model are as follows.
this issue, along with alternatives. A recent review of the Hunter Valley
Operations managed by Rio Tinto Coal Australia Pty Limited (RTCA) is Phase files
used as a case study. RTCA has investments in several mines in the same
area which currently share rail and port facilities with other producers. Each independently scheduled area will be referred to as a phase –
This case study will discuss the many issues involved in these mines and Bengalla has four phases, HVO has 17 and MTW has six. As
show how a coordinated approach can improve the profitability and shown in Figure 1, coal phases are mined vertically, one strip at a
performance of the mines. time. The mining consists of removing waste overburden and then
removing the exposed coal seam. When the bottom of one strip is
INTRODUCTION reached, the operation advances to the next strip. The actual
implementation of this is more complex, allowing for slope
Rio Tinto Coal Australia Pty Limited (RTCA) manages the Coal layback and other equipment issues. The existing computer
& Allied (CAN) operations in the Hunter Valley, consisting of generated equipment schedules, consisting of pit, strip, bench,
Bengalla (BEN), Hunter Valley Operations (HVO) and Mount block and start time, were used to create, for each phase, a
Thorley Warkworth (MTW). The mines produce a range of coal time-based sequence of waste bcm, coal tonnage, ash and other
products for both the export and domestic markets, including qualitative data for use with the computer model.
thermal coal, semi-soft coking coal and low, medium and high
ash export coals. All export coal from these mines and other
producers in the region is transported by rail to the Port of
Newcastle, where it is loaded by Port Waratah Coal Services 5
overburden
onto ships. The rail and port facilities are currently capacity coal
interburden 1 6
constrained. coal
RTCA initiated a review of the way they do business that 2
would allow them to consider all aspects of their operations. 7
They proposed a three pronged attack. The first stage was to Benches
build a computer model that would produce a life-of-mine 3
schedule, then validate it against the MTW operations to ensure
that they could replicate the existing net present value (NPV).
MTW was chosen as it had the most complex operation, with 4
two coal plants and the ability to do one- and two-stage washing ips
and bypass. The second phase was to review all the value drivers Blocks Str
for the operation. The methodology and options explored may be
of interest, and show what can be achieved or uncovered when FIG 1 - Coal seam schematic.
the parameters are opened up for renewed appraisal. This paper
outlines some of the key analyses undertaken by the author in
this work. Other mines may benefit from the same type of Revenue
approach. The policies investigated included:
Coal price varies based on thermal energy and ash content. The
(a) selection of the coal ash/yield set point to maximise ash content also controls product groups:
revenue; • semisoft,
(b) determining the best wash plant to use for each coal • low,
seam, taking yield, processing and haulage costs into
consideration; and • medium and high ash export, and
(c) the application of equipment capital costs when changing • domestic.
optimisation sequence. Modelling these attributes allows the revenue to be calculated
Policies (a) and (c) are discussed in detail below – (b) is not for each coal class.
discussed further. A review of the effect of changes to pit limits,
Costs
1. Principal Consultant, Strategy Optimisation Systems Pty Ltd, Fixed and variable costs were allowed for mining equipment,
66 Rathmullen Quad, Doncaster Vic 3108. mining services, wash plant processing, general overheads, rail
Email: [email protected] haulage and port loading.
C WHARTON
Constraints data. MTW is currently using a set point of 1.6. Table 1 shows
the total coal tonnage available when running the optimisation at
Equipment capacities, mining and wash plant throughputs, rail six other set point settings. As can be seen, the yield drops
and port limits were entered based on existing conditions. dramatically when the set point is lower than 1.5 – price
Practical development constraints related to pit access and lead differentials would have to be significant to make this
times were also included. The optimisation algorithm is worthwhile. Yield increases above 1.6 and it will be the ash and
discussed in the Appendix. qualitative values that determine whether the additional product
can be sold.
Outputs from the model
The key outputs are listed below. TABLE 1
Coal tonnage by wash plant set point.
Period schedule
Set point Domestic Export Total Tonnes diff
Shows policy setting(s), cash flow, quantities of waste by (Mt) (Mt) (Mt) (%)
equipment, coal and product tonnage, ash percentage and other 1.40 7.3 197 204 -26.3
quality attributes, by class, phase and equipment operating hours.
1.45 8.7 229 238 -14.0
Phase summary 1.50 10.1 262 272 -1.9
1.55 12.4 264 277 -0.2
More detailed break-down of scheduled quantities and attributes
by phase by period. 1.60 14.2 263 277 0.0
1.65 15.9 269 285 2.9
Constraints 1.70 18.0 280 298 7.6
Shows what is limiting the schedule for each period.
Investigation of options The ‘best’ set point policy will depend on the coal prices for
each of the coal categories, and plant and product constraints. If
there is little premium for a semi-soft product, then this material
Calibration runs may have more value being blended with low and medium ash
Existing mining plans were used to get the annual wash plant coal grades to achieve better overall value. Figure 2 shows the
tonnage inputs as a target for the calibration fixed sequence life-of-mine total tonnage and average ash content, for each coal
schedule. The computer program was restricted to only find those class, for a fixed and a variable set point run. With the variable
tonnages. Comparison of the calibration run with existing set point run there is less total semi-soft material (-8.8 Mt)
schedules showed physicals within one per cent and NPV within and more low ash (9.1 Mt), high ash (12.1 Mt) and domestic
five per cent. (3.3 Mt). The ash percentage, by coal class, tends to be slightly
higher with the variable set point, however, there is an increase in
coal yield of 15.7 Mt and an improvement in the NPV.
Initial optimisation runs
An optimisation was then run using only the plant capacity Pit limits and extents
constraints to see what other value might be generated. This may
appear unrealistic, but the computer program was run to see which A series of new pit models was produced based on five
pits generated the best NPV and what it would then do with the margin-ranked shells for each of four different coal seam floors
problem. If you don’t relax some constraints, then you run the risk and the existing case. Dragline horizons and configurations were
of getting a similar solution to your fixed schedule, as there is no adjusted, depending on the seam floor depth, and all cases were
freedom or room to move. The optimisation showed a large scheduled in a consistent manner and optimised. All the cases,
improvement in NPV; however, some of the resulting equipment including the current case, were plotted in increasing NPV order
requirements exceeded practical pit capacity, given access and together with the total tonnage, as shown in Figure 3. The current
dump space limitations. What it did highlight was the profitability case is shown as data point A. In general, the shallower options,
of the Battleaxe region by bringing it forward in the schedule. A while they had a smaller equipment footprint, failed to deliver a
further series of optimisation runs were carried out, adding reasonable NPV. The deeper options had more NPV, but also
dragline and shovel equipment limits to each of the pits, overall delivered over twice the current tonnage and excessive mine lives
equipment limits, and revised pit and phase sequence constraints, based on current extraction rates. The work was not in vain as
until a satisfactory schedule was produced. This then became the indicated by two potentially interesting cases (B and C).
new benchmark for other option investigations. A similar situation Case B has an NPV that is almost the same as the current case,
was found while calibrating the HVO mines where the Cheshunt but has a mine tonnage almost 20 per cent less than the current
pit was shown to be a major driver of value. This led to a revision case. Case C has a tonnage that is almost the same, but a higher
of existing plans and the installation of more shovel capacity in the NPV. There might be additional boxcut costs required to integrate
Cheshunt pit. it into the existing design, but the NPV increase is attractive
enough to warrant further investigation. Both these cases should
Ash/yield trade off be scheduled in a detailed manner to confirm their NPV values
and schedules. There is also the opportunity to explore other
Wash plants use a density set point to determine the amount of combinations of pit floors and extents for individual pits based on
ash in the final product. The coal is floated off – the heavier the work done. There may be additional value from going to
material sinks and is rejected. You can obtain better quality different floors in the different pits.
(lower ash) with a lower yield and this may be attractive if the
price increment is sufficient to offset the loss of tonnage. This is Capital application
particularly important when the capacity of rail and port
infrastructure is potentially constraining. To model this sort of The computer optimisation should consider the capital required
complexity you need ash yield data at a variety of set points. if additional trucks or shovels are added to achieve revised
Currently this level of detail is only available for the MTW seam scheduling plans. This leads to more realistic schedules, and
COAL MINING IN THE HUNTER VALLEY – A STRATEGIC MINE PLANNING CASE STUDY
100 30%
90 27%
Mtonnes Product Coal

80 24%
70 21%
Product Ash
60 18%
50 15%
40 12%
30 9%
20 6%
10 3%
0 Exp Dom
0%
SS 1 SS 2 LAE 3 LAE 4 LAE 5 HAE 6 HAE 7 Dom 8
Byp BDT
Tonnes, Fixed SP 47.4 38.1 76.6 50.9 27.5 8.8 7.7 9.2 7.7 7.0
Tonnes, Variable SP 25.0 51.7 73.1 47.0 44.0 17.6 11.0 9.2 11.0 7.0
Ash, Fixed SP (rhs) 6.6% 8.6% 10.4% 12.5% 14.3% 16.8% 23.5% 11.1% 23.5% 15.0%
Ash, Variable SP (rhs) 6.5% 8.7% 10.5% 12.5% 14.4% 16.6% 24.3% 11.1% 24.3% 15.0%
Tonnes, Fixed SP Tonnes, Variable SP Ash, Fixed SP (rhs) Ash, Variable SP (rhs)
Product codes
SS semi-soft, LAE low ash export, HAE high ash export, DOM domestic
FIG 2 - Ash/yield – fixed versus variable set point.
160% 12000.0
140% NPV
10000.0
Tonnes
Current C
120%
B A 8000.0
Tonnes (m)
100%
NPV (%)
80% 6000.0
60%
4000.0
40%
2000.0
20%
0% 0.0
W 2 L 4. 2 65 6
C 1 1 18 T 26 3
.3 9 0
3
L .6 7 1
N 5 3 7 8
FJ t 8 4 2
M 16 3 T 37
B 2 L T 42
3 16 T 35
F 3 8 T 5
W U R 3 L 8 .2 T 6 9
B 2 L .1 5 7 6
C W 8 8 T 5
D G M D 1 1 7 .7 1 8 6
4 4 1 3 7
7
4 D 3 .7 1 6
W B F 2 L 3 .2 2 7
W W K 4 35 . T 5 3
C K 4 15. T 5 14
B Q 418 33. 9 17
11 0
Q 19 L 9 T 958
W Q 4 L 7. 60 2
U 5 0 3 7
Q 7 18 T 37
M M c 14 T 2
U 9 22 T 9
C K 17 37 T 44
B 1 45 7 T 989
C CU 010 L 2 3 T 91
55 T 00
W 1 L 3 1 T 23
0 45 T 65
rB 01 L 2 .1 T 860
B flee L 1 9 T 12
B 4 L 2 .6 8 3
T 16
C 4 1 L 2 .4 8 2 5
W 0 2 T 84
W U R 2 L .9 69
G 8 L T 3
2 2 L T 53
29
4
B re L 2 .8 T 9 5
N 1 3 T 4
C 0 L 1 .6 T 4 3 3
2 M 1 L8 T 2
6 2 T 0
G G t + L .9 1 6 6
C J 2 t L .8 3 8 4
4
B NQ 2 L .5 45
u r K 1 .6 1 2
G 8 L 1 . 5 43
ur 61 8 9 7
B 6 L .2 3
R L .9 0
FJ L 22 3
. 4
4
02 L 6 6
D D a p .2 3
4
C 6 L 5 .4 T
T
5
D 6 3
3
Q L .
F n 1
K 5 0
G J8 1 L
K 5 9
fle 4 L
B D8
W J
W 6
N 2
M
R
ur
F
ur R
D
J
M
G
N
u
M
G
FIG 3 - Alternative pit configurations.
stops equipment from being added and not adequately used. measured and assessed on a common basis. The model has also
While the NPV improvements may not be large, they are easy to been used to investigate plant expansions to cope with new coal
implement and have little impact on the running time. The NPV fields, and decisions on plant location, including truck haulage
will also be more realistic and agree more closely with and/or conveyor options.
traditional assessments.
Long life projects
Hunter Valley integrated models
The HVO mine has a long projected life which poses issues for
Once each mine site is calibrated it can be added into an evaluations based solely on NPV. This problem is not unique to
integrated model. This model maintains individual site mine, coal, but applies to all mines with long lives. If you have several
plant and equipment limits, and adds potential port and rail alternatives with similar NPVs, what other criteria can you apply
constraints. This means that alternative expansion plans can be to determine which is ‘better’? The reason for this problem is
C WHARTON
that the discount decay means that there is little contribution to Carbon tax issues
NPV once the mine life gets over 18 to 20 years. Figure 4 may
help to illustrate the problem. It shows the yearly discount factor Greenhouse gas emissions are created directly from exposed coal
for a ten per cent discount rate. The contributions at years ten, 20 seams and indirectly by the use of diesel and electricity during
and 30 are approximately 40 per cent, 15 per cent and five per the mining process. The coal products also create greenhouse gas
cent, respectively. If we assumed that the net revenues were emissions when used by customers for power generation and
evenly distributed over the mine life, then the first five years converting iron ore into steel. Garnaut (2008) has stated:
contributes approximately 40 - 45 per cent of the NPV, and by If the coal industry is to have a long-term future
year ten you have accounted for 60 - 70 per cent of the value. in a low-emissions economy, then it will have to
be transformed to near-zero emissions, from
10% Discount Rate
source to end use, by mid-century.
1 100%
0.9
RTCA has a published climate change action plan and is
currently supporting research into sequestration through their
Per cent contribution

0.8 80%
support of the CO2CRC Otway carbon capture project, and into
Discount factor
0.7
0.6 60% technologies which reduce carbon dioxide emissions through
0.5 their support of the Coal21 fund.
0.4 40%
The computer models are capable of considering a range of
0.3
0.2 20%
future carbon priced scenarios. It is straightforward to add
0.1 another cost or range of costs into the optimisation cashflow
0 0% logic and see the financial impact of a carbon tax on the project
0 10 20 30 40 50 value, but there are several other questions that need to be asked:
Discount 10yr 20yr 30yr 40yr 50yr What is the appropriate impost? There have been several
suggested rates but no firm government policy as yet. How will it
FIG 4 - Discount rate and value contribution by period. escalate with time? You can run a series of cases to understand
the possible impact. Will any of the costs be passed on to the
customer? If they cannot, then it is important to determine where
There are other qualitative measures that should also be the company is positioned in the industry cost curve, and what
assessed over the mine life: the likely impact is on their viability, compared with their
• plant throughput and head grades, competitors. Are customers likely to look for a cleaner substitute,
and does this have an impact on proposed mine life? Finally, are
• peak mining rates and equipment use, there technologies available or being developed that may assist in
• capital requirements, and mitigating the carbon footprint, and at what cost and timeframe?
• return on investment and payback periods.
CONCLUSIONS
You could also look at other metrics, such as capital intensity, or
product produced for different periods of the project, then choose Reviewing and re-evaluating the value drivers for the project
the case that had better values at the start of the project. A similar turned out to be a worthwhile exercise. It did require a lot of
approach is to break the mine down into Business Risk Periods work setting up the data to support the analyses, but RTCA
(BRP) (Steffen, 1997). The BRP is a period, usually seven or eight gained insights into what was producing value and where they
years, over which the mine makes investment decisions and could focus their current and future efforts. Each of the mines
expects to make a profit. It distinguishes risks associated with benefitted from a review of the optimisation possibilities. The
business from those associated with geology. You break the mine
runs revealed alternative sequencing arrangements or situations
life into BRPs and plan within those BRPs. Project risks change
where additional equipment added value. With a computer model
over time – this can be highlighted in different BRPs as shown in
Table 2, which is the author’s illustrative interpretation of the way in place, RTCA gained the ability to analyse a lot of alternatives
in which risk may change through the life-of-mine and the BRPs. in a speedy and consistent manner. This has helped them make
For example, price is not normally an issue at the start of the informed decisions regarding future expansion plans.
project, as you may have contracts, hedging, or recent history to
guide you. Costs are reasonably well known and you should have ACKNOWLEDGEMENTS
measured reserves to start with, so the geology should be low risk.
The more likely risks for start-up mines could be process The author would like to thank RTCA for allowing the
technology, initial plant performance or ramp up. For long life publication of this paper. M Scott, the general manager of mine
mines, there might be the consideration of new technology or planning, has been particularly supportive through all the
product substitution. computer modelling.
TABLE 2
Business risk periods.
Component BRP 1 BRP 2 BRP n

Equipment/technology • Is the equipment proven? • New competitors • Can you take advantage of new
• Can you start small and ramp up? • Position on cost curve technology?
Price • Not normally an issue as the grades • Are prices rising or falling? • Critical for incremental costs in the
are better in the inner pits outer pit shells
Cut-off • Can you use variable cut-offs? • Are there any stripping hurdles? • Can you keep the mill full?
• Can you access the orebody?
Payback • Payback on initial investment • Payback on next investment • Payback on next investment
COAL MINING IN THE HUNTER VALLEY – A STRATEGIC MINE PLANNING CASE STUDY
REFERENCES The formulation of the problem, using the Lane nomenclature,

is given by the equation:
Australian Coal Association, 2004. Coal 21 National Action Plan
[online]. Available from: <http://www.coal21.com.au/Media/COAL
Max ⎧ V '( R − r, T + t ) ⎫
21_Action_Plan.pdf>. V '( R, T ) = 0≤ r ≤ R ⎨c( t, r, w ) + ⎬
Cooperative Research Centre for Greenhouse Gas Technologies all ω ⎩ (1 + δ )t ⎭
(Co2CRC), 2008. 2008 Annual Report [online]. Available from:
<http://www.co2crc.com.au/dls/annreps/08/ResearchPrograms_CO2
CRC_Annual0708.pdf>. where:
Garnaut, R, 2008. Garnaut Climate Change Review Final Report – 30 the maximum NPV (V) of a resource (R) at a time (T) is
September 2008 (full version) [online]. Available from: <http://www. calculated by finding the best strategy (ω) that maximises the
garnautreview.org.au/> [Accessed: 27 October 2008]. sum of the cash flow (c) for an increment of the resource (r) and
Lane, K F, 1988. The Economic Definition of Ore: Cut-off Grades in the maximum NPV of the remaining reserve.
Theory and Practice (Mining Journal Books Limited). The remaining resource cash flows are discounted (by δ) and
Rio Tinto Coal Australia Pty Limited, 2006. RTCA Climate Change start after the increment of resource is mined (at time t). In this
Action Plan October 2006 [online] Available from: <http://www. formulation the maximum value (V) of the entire reserve and
riotintocoalaustralia.com.au/documents/2_-_Climate_Change_Action
_Plan.pdf>. remaining reserve is replaced by an estimate of the maximum
Rio Tinto Coal Australia Pty Limited 2007. 2007 RTCA climate change
value (V’). The phase files define the mining sequence
review [online]. Available from: <http://www.riotintocoalaustralia. increments. The policies will control what actions to take in each
com.au/documents/2007Review_CCAP.pdf>. period, the revenue and costs will determine cash flow and the
Steffen, O, 1997. Planning of open pit mines on a risk basis, The Journal constraints will control how much of the resource is mined.
of the South African Institute of Mining and Metallurgy, March/April The resource is mined from the beginning, and an estimate of
97:47-56. the remaining value is determined by scheduling the resource,
Wooller, R, 2007. Optimising multiple operating policies for exploiting using default properties, in an initial pass. These defaults may be
complex resources – An overview of the COMET scheduler, in a breakeven cut-off grade, or in the case of coal, an assumption
R Dimitrakopoulos), pp 289-296 (The Australasian Institute of Mining that all uncovered coal will be used with a yield based on the
and Metallurgy: Melbourne). current wash plant settings. The algorithm then repeats the
calculations, optimising the operating policies using the previous
value estimate, until the requested number of iterations is
APPENDIX
reached or no improvement in NPV is found compared with the
previous iteration. Typically, this process requires less than ten
Optimisation algorithm iterations to converge on a solution. Wooller (2007) provides a
The optimisation algorithm uses dynamic programming with good description of all the inputs and operation of the computer
iterative successive approximations to search for the maximum program. This algorithm uses a number of sequences in ‘phase
NPV over the life of the mine, using defined project policies. The files’. These are well suited to coal since within a ramp area a set
policies currently include, but are not limited to, the sequencing of strips must normally be mined in order. Areas with different
of the phases, the selection of the wash plant set point and the ramp access often have several areas that can be advanced (or
application of equipment capital costs. There is no guarantee that delayed). The policies will control what actions to take in each
the solution will converge on the optimum, however, this period in order to maximise the NPV as the resources are mined.
technique has been used successfully to improve project values.
HOME
A New Methodology for Flexible Mine Design

B Groeneveld1,2, E Topal3 and B Leenders4
ABSTRACT maximise the equipment utilisation and meet mill requirements

Uncertainty and risk are invariably embedded in every mining project.
while matching the mining rates previously derived by the
Mining companies endeavouring to maximise their return for optimisation. The last stage generates a single mining sequence
shareholders make important strategic decisions which take years or even from alternative sequences produced in the third stage by using a
decades to ‘play out’. Therefore, developing a model that analyses the new algorithm based on the simulated annealing method. Leite
potential payoff of a decision based on current fixed assumptions is and Dimitrakopoulos (2007) examined a stochastic optimisation
severely flawed. A model that incorporates uncertainty and is able to model based on simulated annealing algorithm for open pit
adapt will help deliver a design with a better risk-return profile. In this mines and applied it to a copper deposit for risk analysis. The
paper, a new approach is developed in order to have a design that is
study shows the stochastic approach generates 26 per cent higher
flexible and able to adapt with change. This is achieved by developing a
mixed integer programming model that determines the optimal design for net present value (NPV) than the conventional scheduler. Also,
simulated stochastic parameters. This research has incorporated the study suggests that life-of-mine schedules which incorporate
optionality (flexibility) in relation to mining, stockpiling, processing plant geological uncertainty lead to more informed investment decisions
and port capacity. The results are promising and are helping decision and improved mining practices. Leite and Dimitrakopoulos (2010,
makers to think in terms of value, risk and frequency of execution. in this volume) and Ramazan and Dimtrakopoulos (2007) present
a stochastic integer programming (SIP) formulation for mine
INTRODUCTION production scheduling. The approach explicitly integrates the
grade uncertainty in the scheduling problem to generate a more
Mining projects are characterised as being highly uncertain and risk-robust solution. The same copper deposit has been utilised in
variable mainly due to the volatile nature of commodity prices Leite and Dimitrakopoulos (2007) for application and the
and uncertainty around geological conditions encountered in
stochastic formulation is shown to produce 29 per cent higher
orebodies. Uncertainty can arise from many different sources
including market prices, grade distribution, ground conditions, NPV than the schedule obtained from conventional scheduling.
equipment reliability, recovery of ore, human, capital and Meagher, Sabour and Dimitrakopoulos (2010, in this volume)
legislative change (Topal, 2008). The mining industry will be introduced an approach to integrate the block destination
more sustainable if projects are developed in a manner that flexibility in the process of assigning a value to mining blocks in
increases the flexibility to respond to uncertainties in the planning process via real options valuation (ROV) considering
business cycle. For example, the global minerals industry has geological and market uncertainties. The proposed approach;
seen an unprecedented demand for its products in recent years, firstly assigns a dollar value to each mining block considering
however the industry has struggled to change its level of supply the different aspect of uncertainty and management flexibility.
in response to price movements. Being able to design an Secondly, it utilises a minimum cut algorithm for designing a
operation that has the flexibility to respond to this change lower risk long-term mine plan. Application to a case study
quickly will deliver better returns to stakeholders. demonstrates significant differences in block value estimates.
Geological uncertainty and risk have been incorporated in A developing decision-making tool aimed at increasing the
optimum mine planning and design by a few studies to date. flexibility of an engineering system is real options ‘in’ projects.
Ramazan and Dimitrakopoulos (2004) developed a stochastic Significant research into this method has been undertaken by de
based mixed integer programming (MIP) model for multiple Neufvillie and his colleagues with applications in various
elements that uses several simulated orebodies in order to industries. This method is located midway between financial real
minimise the grade uncertainty in the life of the mine schedule. options analysis (which does not deal with system flexibility) and
This model also takes into account risk quantification, equipment traditional engineering approaches (which does not deal with
access and mobility, and other operational requirement such as financial flexibility). A popular example used to explore the
blending, mill capacity and mine production capacity. Godoy and concept of real options ‘in’ projects is that of a multi-story car
Dimitrakopolus (2004) developed a new set of ways to generate a park. Flexibility in this situation is in the design of the footing and
mine production schedule under geological uncertainty. The first columns of the building so that additional levels can be added at a
stage of the method generates a stable solution domain which later date. This flexibility comes at a cost, and the designer must
shows the possible ore and waste extraction rates for a given determine if this is warranted. An example of the opportunity this
open pit. The second stage generates optimum ore production technique poses for mining is the Chilean mine in the ‘Cluster
and waste removal under uncertainty. The third stage generates a Toki’ region, and was undertaken by de Neufville , in conjunction
series of physical schedules which obey slope constraints, with Codelco. In this example, a staged development of the real
option ‘in’ projects methodology is used where different operating
1. GAusIMM, PhD Candidate, Mining Engineering Department, plans are designed to respond to changing prices. Truck fleet
Western Australian School of Mines, Curtin University of Technology, capacity and crusher size were altered in the different operating
GPO Box U1987, Perth WA 6845. plans. The application of this method resulted in approximately 30
Email: [email protected]
to 50 per cent more accurate project value than current estimates.
2. Newcrest Graduate Mining Engineer, Australia. This approach provides a strong basis on which to grow real
3. MAusIMM, Associate Proffesor and Head of Mining Engineering options ‘in’ projects theory for mining. However, there are several
Department, Western Australian School of Mines, Curtin University of deficiencies in the current model. First, the initial scenario
Technology, GPO Box U1987, Perth WA 6845. construction used in the model limits the flexibility up front in the
Email: [email protected] model and prevents the optimal design from being chosen.
4. MAusIMM, Principal Advisor – Strategic Development, Rio Tinto Therefore, how useful is this technique for valuing flexibility?
Iron Ore, Expansion Projects, Resource Development Division, Perth Secondly, the model fails to deal with variations in grade and
WA 6000. Email: [email protected] recovery in a transparent manner; one of the key drivers. Finally,
B GROENEVELD, E TOPAL and B LEENDERS
the model does not incorporate options at all stages of a typical Overview of resource model
mine value chain (de Neufville, Scholtes and Wang, 2005; Wang
Resource characteristics are a driving force in mine design. The
and de Neufville, 2005, 2006; Cardin, 2007; Cardin, de Neufville
MIP model uses a resource model to provide a representation of
and Kazakidis, 2008). This paper outlines a new methodology to material that is available for processing through the life-of-mine
evaluate the flexibility of strategic mine design under uncertainty, (both ore and waste is considered). The representation of the
using mixed integer programming (MIP) and Monte Carlo resource is carried out by parcels of material. A parcel of
simulation (MCS). An application of this methodology to a material can be defined as a quantity of material with an average
hypothetical case study will be undertaken in order to show the grade determined by the weighted average of the grade bins
power of the model to handle complex strategic decisions. contained within the parcel. A parcel may be made up of one or
more grade bins. A grade bin represents a quantity of material at
METHODOLOGY a specified grade. This is incorporated to provide a higher level
of detail to the model which will alter the decisions on how
In order to evaluate the flexibility in strategic mine design, this material is processed, whilst minimising the number of integer
research employs MIP and MCS. In particular, MIP allows for ‘go’ variables. These parcels are designed to represent a physical
or ‘no go’ decisions to be modelled for optimal execution under a constraint on the resource. The most common physical constraint
set of uncertainties. Uncertainties (or stochastic parameters) can be is the vertical mining constraint, which is included in the model
simulated using MCS. In this way, each model (or trial) represents through parcel dependency. Mining of the grade bins within a
a single path of a lattice tree (or binomial tree). parcel can occur in any order, as long as the average parcel grade
(within a nominal deviation) is extracted each period. This forces
the model to take waste and ore in the same proportion.
Description of model components
Overview of the design options
The model consists of three main components which feed the
MIP model; resource model, design options and stochastic Flexibility is included through various design options in the MIP
parameters (Figure 1). Running the model multiple times model. Solving the MIP models will determine which options are
generates a database of optimal designs for a given ‘state-of-the- executed and when. A full set of design options is dynamically
world’. This data set then provides a pathway to determine the incorporated in the model which determines if and when these
flexibilities that provide the best risk-return profile. options should be executed. These options are broken into four
categories: mine, preprocessing stockpiling, processing plants
and port capacity. More than one option type can be executed in
each period, hence these are not mutually exclusive decisions.
An illustration of the material flow and points where design
options may occur is shown in Figure 2. Some assumptions have
been made to simplify the model at this early stage of
development. These assumptions can be removed with further
refinement to the model; one type of circuit exists in each plant
with one set of beneficiation characteristics; port stockpiles are
not available in the model. This means the model must ship
material as soon as it is processed.
Available mine options in the model

Mine options are incorporated in the model to reflect mining
capacity constraints that exist in an operation. It is not feasible to
have unlimited mining capacity, due to the high capital cost
associated with additional capacity and/or technical pit constraints
(geotechnical and equipment interaction). Mine options can be
FIG 1 - Conceptual diagram of how the various model components modelled to reflect truck capacity or shovel capacity. This type
feed the mixed integer programming model and the result set. of decision is repeatable many times in each period (ie you can
FIG 2 - Material flow in the model and the location of the design options.
A NEW METHODOLOGY FOR FLEXIBLE MINE DESIGN
purchase more than one truck of the same type), thus mine Outline of what stochastic parameters are included
options are represented as integers. This allows for one or more
trucks of the same type to be purchased in each period. Uncertainty in the mining process is incorporated through the
market price, cost (capital and operating), utilisation of
Available stockpile options in the model equipment, plant recovery and time to build an option. Values for
these various inputs are simulated through a MCS process.
Stockpiling is used in mine operations for many reasons including
blending of material, storage of excess mine production and
storage of low-grade ore for future production. Long-term Model formulation
stockpiling is included in the model, allowing material to be stored
on a stockpile in time (t) and removed in subsequent periods The developed MIP model optimises the available mine,
(t+1…t+N). A further ability of the stockpile option is its ability to stockpile, plant and port flexibility for a simulated scenario.
represent long-term waste dumps. This functionality allows the These various design options dictate how the system is
model to consider waste movement, and dynamically changes the configured and consequently the amount of production that can
cut-off grade. Waste dumps are developed by entering an option occur. They also dictate the financial viability of the operation
that is similar to a stockpile but has no plants for the material to and drive both revenue and operating costs. An outline of the
flow to, forcing it to remain on the waste dump. mathematical formulation is provided below.
Available plant options in the model

Objective function
Plant flexibility is incorporated to model the options managers
have around processing of mined ore through varying plant The objective function seeks to maximise before tax net present
designs. A plant option is characterised by its capacity, capital value (NPV):
cost, fixed operating cost, recovery characteristics and grade
T L M S L O
limits. A processing plant is the link between the mine and the 1
port in the flow of material in the model. Plant options can also be ∑ (1 + r ) [ S ∑ M
* t t l,t − ∑ C m I m , t − ∑ C sY s , t − ∑ C lYl , t − ∑ C oYo , t
t=1 l=1 m=1 s=1 l=1 o= 1
dependent on other plant options being built, allowing the idea of M L O M L
modular plant capacity to be modelled. That is where a high initial − ∑ Dm IDm , t − ∑ Dl IDl , t − ∑ Do IDo , t − ∑V m, t X m , t − ∑ Vl , t X l , t
m=1 l=1 o= 1 m=1 l=1
capital cost is incurred to allow for expansion at a later date with a
smaller capital cost.
S O M L O
− ∑ Vs , t X s , t − ∑ Vo , t X o , t − ∑ Fm , t I m , t − ∑ Fl , tYl , t − ∑ Fo , tYo , t
Available port options in the model s=1 o= 1 m=1 l=1 o= 1
Port options allow the sale of material to customers in the model.

M L O L P, B
Enough port capacity must exist in a period in order for any plant
production to occur in that period. For example, if we have two ∑ FR m, t IDm , t − ∑ FRl , t IDl , t − ∑ FRo , t IDo , t − ∑ L l ∑X p, b , l , t
m=1 l=1 o= 1 l=1 p= 1, b = 1
million tonnes of plant capacity and no port capacity in the first
period, then the production from the plant is forced to be zero. S P, B, L
−∑ L s ∑ XI ⎤
Overview of how mine scheduling works in the model s=1 p = 1 , b = 1 , l = 1 s ∈l
p,b,s,l,t ⎦⎥
Resource characteristics are a driving force in mine design. The

MIP model uses a resource model to provide a representation of where:
material that is available for processing through the life of mine
(both ore and waste is considered). r* is the rate of return on the project
The representation of the resource is carried out by parcels of St is sale price in time period t (in $/metal unit)
material. A parcel of material can be defined as a quantity of Ml,t is the metal units exiting plant l in time t
material with an average grade determined by the weighted
average of the grade bins contained within the parcel. A parcel Cm, Cs, Cl, Co is the capital cost of mine m or stockpile s or
may be made up of one or more grade bins. A grade bin represents plant l or port o
a quantity of material at a specified grade. This is incorporated to Im,t is the execution integer on mine option m in
provide a higher level of detail to the model which will alter the time t
decisions on how material is processed, whilst minimising the
number of integer variables. These parcels are designed to 0 is no execution of option;
represent a physical constraint on the resource. The most otherwise, is the number of times the option
common physical constraint is the vertical mining constraint is executed in time t.
which is included in the model through parcel dependency.
Ys,t,Yl,t,Yo,t is the execution binary on stockpile s or plant
Mining of the grade bins within a parcel can occur in any order l or port o
as long as the average parcel grade (within a nominal deviation)
is extracted each period. This forces the model to take waste and 0 is no execution of option;
ore in the same proportion. otherwise, the option is executed in time t.
Options in the model to test best plant or stockpile Dm, Ds, Dl, Do is the disposal cost of mine m or stockpile s
location through varying mining costs or plant l or port o
The optimal location for a processing plant varies with time as IDm,t is the disposal integer on mine option m in
the resource is mined in different regions. Therefore, determining time t
the best location for a plant or a stockpile is not a simple case 0 is no disposal occurs in time t;
and must consider these multiple uncertainties as it will most
otherwise, is the number of options disposed
likely change over the life of the project. Different plant
of in time t.
locations and stockpile locations can be tested by developing a
mining cost which varies by parcel and destination in the model. IDl,t ,IDo,t is the disposal integer on plant l or port o
B, t t
0 is no disposal occurs in time t;
∑ X p + 1 , b , tt ≤ R p + 1 ∗ ∑ Y p , tt ∀p, t
otherwise, is option is disposed in time t. b = 1 , tt = 1 tt = 1
Vm,t, Vs,t, Vl,t, Vo,t is the variable cost of mining a tonne of ore where:
from mine m or stockpile s or plant l or port Rp+1 is the resource of the successor parcel p + 1
o in time t
Xp+1,b,tt is the tonnage mined from the successor parcel p + 1 bin
Xm,t, Xs,t, Xl,t, Xo,t is the tonnage processed through mine m or b in time tt
stockpile s or plant l or port o in time t
Sequencing constraint 3
Fm,t,Fl,t,Fo,t is the fixed cost of mining from mine m or
plant l or port o in time t This constraint is a set packing constraint that forces a parcel to
only be fully mined once:
FRm,t,FRl,t,FRo,t is the reduction in fixed cost of disposing of
T
an option in time t
∑Y p, t ≤1 ∀p
Ll is the cost of mining a tonne of ore to plant l t=1
Ls is the cost of mining a tonne of ore to Equal mining of a parcel

stockpile s
This constraint forces the model to take high-grade ore, waste and
Xp,b,l,t is the tonnage mined from parcel p bin b to low-grade ore in equal proportions. This prevents the model taking
plant l in time t high-grade in the first period, followed by low-grade in the second
XIp,b,s,l,t is the tonnage mined from parcel p bin b to period and waste in the following period. A minimal deviation (γ)
stockpile s to plant l in time t of two per cent was allowed to prevent an infeasible solution:
The objective function represents the following: B
⎡ B ⎤
∑X p, b , t G p , b ≥ ⎢∑ X p , b , t ⎥G p ∗(1 − γ%) ∀p, t
The revenue from the sale of the ore less the b =1 ⎣b = 1 ⎦
capital cost of building an option less the B
⎡ B ⎤
disposal cost of reducing capacity less the
variable cost of processing ore less the fixed cost ∑X p, b , t G p , b ≤ ⎢∑ X p , b , t ⎥G p ∗(1 + γ%) ∀p, t
b =1 ⎣b = 1 ⎦
of maintaining an option; all multiplied by the
relevant discount factor for the cash flow in time t. where:
The model seeks to maximise this relationship. Gp,b is the grade of parcel p and bin b
The constraints in the model can be divided into five Gp is the grade of parcel p
categories: production, mining, stockpiling, processing plant and
port constraints. Flow balance constraint
This constraint links the flow paths in the model and ensures that
Production constraints the material available to the processing plant and stockpiling
options originates from the resource:
Resource constraint L S,K,L
X p, b , t = ∑ X p, b , l , t + ∑ XI p , b , s , k , l , t ∀p, b, t
This constraint makes sure the total amount of material extracted l=1 s = 1 , k = 1 , l = 1 s ∈l
from a mining pit has an upper bound based on the resource.
This constraint is applied at a parcel and bin level in the model: Mining constraints
T
∑X p, b , t −R p , b ≤ 0 ∀p, b Mining requirements

t=1
The constraint makes sure that mining includes all movement to
where: plant options, stockpile options and movement off stockpiles:
Xp,b,t is the tonnage mined from parcel p bin b in M P, B S,L
time t ∑X m, t = ∑ X p, b , t + ∑ XOs , l , t ∀t
m=1 p= 1, b = 1 s = 1 , l = 1 s ∈l
Rp,b is the resource of parcel p bin b
where:
Sequencing constraint 1
XOs,l,t is the tonnage sent from stockpile s to location l in time t
This constraint in conjunction with the next constraint forces the
binary value to be one in the period the parcel is fully mined. Mining capacity limit
This then allows the model to mine any successor parcels of ore:
B, t
This constraint ensures that mining only occurs if there is
∑ X p , b , tt ≥ R p ∗Y p , t ∀p, t sufficient capacity in a period to handle the movement. Capacity is
determined dynamically based on when mining options are
b = 1 , tt = 1
executed:
where: t t
Rp is the resource of parcel p ∑A I

m, u , t m, u − ∑ Am , u , t IDm , u ≤ X m , t ∀m, t
u =1 u =2
Sequencing constraint 2 where:

This constraint ensures that a parcel’s predecessor is mined Am,u,t is the capacity of mine option m that was executed in
before the successor is mined: period u in time t
Mine option disposal constraint Stockpile capacity constraint

This constraint ensures disposal of an option can only occur if This constraint makes sure the tonnage of material stockpiled
the option has been built. For example, if a mine option is built in across all plant locations does not exceed the stockpile capacity:
period one, and in period five there is no more material to mine,
L,t L,t t
then the model can dispose of this capacity in order to reduce the
fixed cost incurred: ∑ XI s , l , tt − ∑ XOs , l , tt ≤ ∑ A sY s , tt ∀s , t
l = 1 , tt = 1 s ∈l l = 1 , tt = 2 s ∈l tt = 1
t
IDm , t ≤ ∑ I m,tt − 1 ∀m, t where:
tt = 2
XOs,l,t is the total tonnage sent from stockpile s to plant l in
t t
time t
∑ ID m , tt ≤ ∑ I m,tt − 1 ∀m, t
As is the total capacity of stockpiles
tt = 2 tt = 2
Maximum execution Stockpile grade constraint on bins

Since the mine option can be modelled as an integer variable, This constraint applies the grade limits of the stockpile bins to
limits on how many times it can be applied in the model may be material entering each stockpile bin:
included. This is an optional constraint which can be turned on or t t
off when running the model. ∑ G p , b XI p , b , s , k , l , tt ≤ ∑ GU s , k XI p , b , s , k , l , tt ∀p, b, s, k, l, t
tt = 1 tt = 1
t t
Period constraint
∑ G p , b XI p , b , s , k , l , tt ≥ ∑ GL s , k XI p , b , s , k , l , tt ∀p, b, s, k, l, t
This constraint restricts the number of mine options built in a tt = 1 tt = 1
period to the period constraint maximum:
where:
I m,t ≤ PGm ∀m, t GUs,k is the upper grade limit of stockpile s bin k
GLs,k is the lower grade limit of stockpile s bin k
Overall constraint
This constraint restricts the total number of mine options built Bin removal constraint
over the life of a mine:
The constraint ensures material moved from the stockpile has
T been added to the stockpile at least one period ago, and that
∑I m, t ≤ OC m ∀m material removed from the stockpile is not removed again:
t=1
t t
where: XOs , k , l , t ≤ ∑ XI s , k , l , tt + ∑ XOs , k , l , tt − 1 ∀s, k, l, t

tt = 1 tt = 2
PC is the period constraint limit or the maximum number of
times an option can be executed in any period where:
OC is the overall constraint which is the maximum number of XOs,k,l,t is the tonnage removed from stockpile s bin k to plant l
times an option can be executed over the life of the project in time t
Stockpiling constraints Bin extraction

This constraint makes the total tonnage of material that is
Total inflow constraint extracted from each stockpile bin equal the overall extraction
from the stockpile:
This constraint makes the total amount of material that is entering K
a stockpile equal the material entering each stockpile bin: ∑ XOs , k , l , t = XOs , l , t ∀s, l, t s ∈ l
k= 1
K
XI s , l , t = ∑ XI s , k , l , t ∀s, l, t s ∈ l
k= 1 Metal extraction
where: This constraint ensures the total amount of metal units extracted
from a stockpile equals the metal units extracted from the
XIs,l,t is the total tonnage sent into stockpile s at plant l in time t
individual grade bins:
K
Flow balance constraint MOs , l , t = ∑ GA s , k XOs , k , l , t ∀s, l, t s ∈ l
k= 1
This constraint restricts the total amount of material coming into
each bin in the stockpile to be equal to the material sent from the where:
each parcel to the bin:
MOs,l,t is the metal units removed from stockpile s at location l
P, B
in time t
XI s , k , l , t = ∑ XI p , b , s , k , l , t ∀s, k, l, t s ∈ l
p= 1, b = 1 GAs,k is the average grade of stockpile s grade bin k
where: Opening limit
XIs,k,l,t is the tonnage sent into stockpile s grade bin k at
plant l in time t This constraint ensures that a stockpile can only be opened once:
T
XIp,b,s,k,l,t is the tonnage from parcel p bin b sent into stockpile ∑ Y s , t ≤ 1 ∀s
s grade bin k at plant l in time t t=1
S
Processing plant constraints
∑ E l , tGM l , t MOs , l , t ∀l, t
s = 1 s ∈l
Grade limits (upper and lower)
where:
This constraint applies the grade limits on a given plant in each
time period. This ensures every plant processes material it can GMl,t is the grade multiple of plant option l in time t
handle:
S P, B Plant option dependency
∑ MOs , l , t + ∑ G p, b X p, b , l , t
Plant option dependency dictates the relationships that occur
s = 1 s ∈l p= 1, b = 1
S P, B
between options. Two types of relationships are available;
≤ GUl ∑ XO s,l,t + GUl ∑ X p , b , l , t ∀l, t one-for-one and one-for-many.
s = 1 s ∈l p= 1, b = 1
One-for-one relationship
S P, B S
∑ MOs , l , t + ∑ G p , b X p , b , l , t ≥ GLl ∑ XO s,l,t This constraint makes sure a successor option is built prior to the
s = 1 s ∈l p= 1, b = 1 s = 1 s ∈l predecessor option being built in an equal ratio. For example,
P, B this can be used to model a modular plant design where an initial
+ GLl ∑ X p, b , l , t ∀l, t investment can be made in plant capacity that has the ability to
be expanded easily for a lower capital than if the initial investment
p= 1, b = 1
was not made (this later expansion is optional):
where: t t − DT
GLl is the lower grade limit of plant l ∑Y l , tt ≤ ∑ Yc , tt ∀l, c, t c ∈ l

tt = 1 tt = 1
GUl is the upper grade limit of plant l
One-for-many relationships
Plant capacity constraint This constraint allows a successor option to be built if its
The constraint ensures that the total tonnage of material processed predecessor option has been built at least once. This can be used
in a period shall be less than the capacity of plant options built and to model a rail link to a plant location where an initial capital
disposed: investment is required. However, once this has been built,
S P, B t t
numerous plants can be built at the same location:
∑ XOs , l , t + ∑ X p , b , l , t ≤ ∑ Al , u , tYl , u − ∑ Al , u , t IDl , u ∀l, t t − DT
s = 1 s ∈l p= 1, b = 1 u =1 u =2 Yl , t ≤ ∑ Yc , tt ∀l, c, t c ∈ l
tt = 1
where:
where:
Al,u,t is the capacity of plant option l built in u in time t
c is the predecessor plant option of plant option l
Plant disposal constraint DT is the lead time on the relationship
This constraint ensures that a plant option is only disposed if the Yc,tt is the execution variable of the predecessor plant c of plant
plant has been built in a previous period. This will result in a l in time tt
fixed cost saving, however an additional disposal cost will be
incurred in the objective function: Port constraints
t
YDl , t − ∑ Yl , t ≤ 0 ∀l, t Port production constraint
tt = 1
The constraint ensures the total tonnage of material processed
t t
through all plant options is less than or equal to the total port
∑ YD l , tt − ∑ Yl , tt ≤ 0 ∀l, t
capacity:
tt = 1 tt = 1
L O
Tonnage produced ∑X l,t ≤ ∑ X o, t ∀t

l=1 o= 1
This constraint restricts the tonnage exiting the plant to be equal

to the material entering the plant multiplied by the plant recovery: Capacity constraint
P, B S
The constraint requires the total tonnage of material shipped in a
Xl , t = ∑ E l , t X p, b , l , t + ∑ E l , t XOs , l , t ∀l, t period to be less than the port capacity:
p= 1, b = 1 s = 1 s ∈l
t t
where: X o , t ≤ ∑ Ao , u , tYo , u − ∑ Ao , u , t IDo , u ∀o, t
u =1 u =2
El,t is the recovery of plant l in time t
where:
Metal units produced Ao,u,t is the capacity of port option o built in u time t
This constraint calculates the metal production of a plant option by
multiplying the metal units in to the plant by the recovery and
Disposal constraint 1
grade multiples for the plant. This is used to calculate the revenue Disposal of a port option may only occur if the option has
of the mine: previously been built:
P, B t
Ml,t = ∑E l,t GM l , t X p , b , l , t + IDo , t ≤ ∑ Yo , tt ∀o, t
p= 1, b = 1 tt = 1
t t
haul for pit 1, and location B has the shortest haul for pit 3,
∑ ID o, t ≤ ∑ Yo , tt ∀o,t
whilst pit 2 has an equivalent haul to either location). In order to
tt = 1 tt = 1
process material through a plant at location A, a rail link of 35km
with a capital of $65 M needs to be built. Likewise, at location B,
Port option dependency
a rail link of 10km needs to be built for a capital of $30 M.
Port option dependency may occur in one-for-one or one-for- Finally, in order to process any material through the port, a rail
many relationships. link from the junction of A and B to the coast needs to be built
for a capital of $20 M. The analysis will look at the system
One-for-one relationship configuration over five periods.
This constraint makes sure a successor option is built prior to the
predecessor option being built in an equal ratio:
t t − DT
∑Y o , tt ≤ ∑Y c , tt ∀o, t c ∈ o
tt = 1 tt = 1
One-for-many relationships
This constraint allows a successor option to be built if its
predecessor option has been built at least once. This relationship
can be used to model a rail link that must be built before any port
can be built:
t − DT
Yo , t ≤ ∑Y c , tt ∀o, t c ∈ o
tt = 1
where:
DT is the lead time on the relationship
Yc,tt is the execution variable of the predecessor port c of port o FIG 3 - Conceptual layout of hypothetical mine.
in time tt
Non-negativity, binary and integer restrictions The model inputs

Multiple options were included in the model of this problem as
Non-negativity summarised in Table 1. A full list of the fixed costs, variable
costs and grade constraints is not provided for simplicity
The following variables are restricted to taking on positive values purposes. In order to simulate the different processing plant
as a negative would represent an infeasible situation: locations available to the model, a differential mining cost was
X p , b , t , XI p , b , s , k , l , t ≥ 0 ∀p, b, s, k, l, t used based on the destination of material. A summary of the
different costs associated with each location is outlined in Table 2.
X m , t , X l , t , X o , t ≥ 0 ∀m, l, o, t
XOs , l , t , XOs , k , l , t , XI s , k , l , t XI s , l , t ≥ 0 ∀s, k, l, t TABLE 1
Options available in hypothetical example.
Integers
Type Cost ($M)
The following variables must take on integer values in the model:
Mine options
I m , t , IDo , t , IDm , t , IDl , t ∀m, l, o, t 1 Mt/a unit 3
2 Mt/a unit 4.5
Binaries
Stockpile options
The following variables must take on binary values; integers with
Waste stockpile (500 Mt capacity) 0
an upper bound of one and lower bound of zero:
Low-grade (30 Mt capacity) 0
Yc , tt , Y p , t , Yo , t , Yl , t , Y s , t ∀p, o, l, s, t, c, tt
Plant options
CASE STUDY – OPEN PIT MINE 5 Mt/a fixed (lead 0.5 yr) 50
10 Mt/a fixed (lead 1.5 yr) 92
An application of the methodology was implemented to a
hypothetical mining scenario. The problem is similar in nature to 5 Mt/a flexible (modular, lead 0.5 yr) 75
an iron ore mine, although it could be applied to any open cut Additional 5 Mt/a flexible (modular, lead 0.5 yr) 30
mine. A single mine site is used in this example. Port capacity
10 Mt/a (lead 0.75 yr) 100
The problem 20 Mt/a (lead 1.5 yr) 175
The operation consists of three mining pits (two high-grade and
one low-grade), two plant locations with associated rail
infrastructure, stockpiling and waste storage capabilities at each The stochastic variables
location, and two port options with associated rail requirements In this problem, it was determined that seven stochastic parameters
(Figure 3). In this diagram, the rectangular boxes represent would be included. These were price, recovery, capital cost,
different plant locations (note that location A has the shortest operating cost, and utilisation for mine, plant and port options.
TABLE 2 Results analysis

Differential mining costs to handle different locations in the model.
Based on the input parameters, 200 trials were run, with
Cost ($/t) Pit 1 Pit 2 Pit 3 CPLEX™ used to solve the MIP model. In total it took three
Plant location A 0.2 0.6 1.8 hours to process the model, which was deemed a good solution
Plant location B 2.2 0.6 0.4 time for this model size. The raw data from the results exceeds
four gigabytes. A results analysis process has been developed
Waste location A 0.1 0.3 0.9
which summarises this data. After processing of the model, the
Waste location B 1.1 0.3 0.2 frequency of execution for each options was analysed (Table 3).
Frequency of execution is calculated by dividing the count of the
Choice of underlying distributions was done through discussions number of times an option is executed by the maximum number
with professionals. No detailed analysis of the underlying nature of times it could be executed. Some categories of options (plant,
of the stochastic variables has been carried out, as detailed port, mine) sum to more than one hundred per cent because
research in other papers is available, which was not the primary multiple expansions of that type can occur in the same time
purpose of this paper (Dimitrakopoulos and Abdel Sabour, 2007; period as the options are not mutually exclusive.
Godoy and Dimitrakopoulos, 2004; Lima and Suslick, 2006; From examining Table 3, several conclusions can be
Morley, Snowden and Day, 1999; Topal, 2008). developed. First, it is evident that larger port capacity options
A summary of the values used for each distribution is as should be investigated as the execution frequency is over 50 per
follows: cent for the 20 Mt/a port option (the largest in the model) for
• price follows a lognormal distribution with a mean of $85, three periods. Second, the mine option with 1 Mt/a capacity and
standard deviation of $25 and a correlation of 0.30 between the plant with 5 Mt/a at location B are not valuable options, as
periods; their execution is lower than 20 per cent. Finally, location A is
• recovery follows a triangular distribution with a maximum preferred over location B, as the rail link options which dictate
value of 90 per cent, likely value of 80 per cent, minimum which locations can be used are executed 94 per cent and 45 per
value of 70 per cent and a correlation of 0.05 between periods; cent for A and B respectively. A value at risk graph (VARG)
shows the risk to return relationship. Figure 4 displays the VARG
• capital cost multiple follows a normal distribution with a
mean of 1.08, standard deviation of 0.20 and a correlation of
0.40 between periods;
• operating cost multiple follows a normal distribution with a
mean of 1.03, standard deviation of 0.10 and a correlation of
0.10 between periods;
• mine equipment utilisation follows a triangular distribution
with a maximum value of 95 per cent, likely value of 75 per
cent, minimum value of 60 per cent and a correlation of 0.22
between periods;
• plant utilisation follows a triangular distribution with a
maximum value of 95 per cent, likely value of 80 per cent,
minimum value of 65 per cent and a correlation of 0.21
between periods; and
• port utilisation follows a triangular distribution with a
maximum value of 95 per cent, likely value of 80 per cent,
minimum value of 65 per cent and a correlation of 0.34
between periods. FIG 4 - Value at Risk for case study.
TABLE 3
Frequency of execution for all options in the model.
Period 1 (%) Period 2 (%) Period 3 (%) Period 4 (%) Period 5 (%) All periods (%)
Mine 1 Mt/a 18 4 5 0 0 5
Mine 2 Mt/a 98 32 9 3 1 29
Rail link to A 94 0 0 0 0 19
Rail link to B 45 1 0 0 0 9
Plant 5 Mt/a (A) 69 26 14 6 2 23
Plant 10 Mt/a (A) 66 14 5 1 0 17
Plant 5 Mt/a modular (A) 90 41 13 2 0 29
Additional 5Mt/a modular (A) 0 82 35 16 1 27
Plant 5 Mt/a (B) 19 4 1 1 0 5
Plant 10 Mt/a (B) 44 19 6 2 0 14
Plant 5 Mt/a modular (B) 41 15 4 2 0 12
Additional 5Mt/a modular (B) 0 35 14 3 2 11
Port 10 Mt/a 98 1 0 0 0 2
Port 20 Mt/a 91 84 53 28 8 53
FIG 5 - Example of decision tree for a different case study.
for this example with the base case representing a fixed mine • Handling of grade variability through the use of conditional
design with no optionality. The design chosen for the base case simulation methods will greatly improve the power of the
was based on the 50th percentile design when the model with model (Ramazan and Dimitrakopoulos, 2004).
optionality was run. This design was then fixed in the MIP model
and reprocessed with the same uncertainties. This shows the • Further investigation into appropriate results analysis
outcome of management not changing the operating policy of the techniques is required to fully understand how the primary
mine. The mean NPV of the base case was $702 M, and for the question of flexibility is answered.
case with options was $1298 M, an 85 per cent increase. Further • MIP performance improvement algorithms need to be
to these analysis methods, experimentation is currently underway investigated; these methods may include reducing the
using various data mining techniques. An open source software feasible region with additional constraints, and/or developing
package called Rapid Miner is currently being used. An example a node selection routine for the branch and bound algorithm
(from a different problem set) of the output generated is shown in that exploits some of the nuances in the model.
Figure 5. This example shows a decision tree with the associated • Application of this technique to underground mining is
percentages of times the decision paid off highlighted in the needed to fully capture the options available to mine
Yes/No boxes at the bottom of the nodes. management. In particular, incorporating the process to
optimise the open cut and underground transition point
CONCLUSIONS would be highly beneficial. This would assist in strategic
planning for the entire orebody.
In conclusion, this paper has developed a methodology to evaluate
the strategic mine design flexibility under a stochastic
ACKNOWLEDGEMENTS
environment. The proposed methodology is a unique approach that
allows flexible mine designs to be justified. The decision-maker is The authors wish to thank Rio Tinto Iron Ore for their support of
supported in their choice and refinement of mine design. this project.
Increasing flexibility in mine designs would be advantageous for
responding to changing business conditions across the full REFERENCES
economic cycle. For the sake of comparison, the proposed
Cardin, M A, 2007. Facing reality: Design and management of flexible
methodology has been implemented to a hypothetical mining engineering systems, master’s thesis (unpublished), Engineering
scenario. The results demonstrated that the value of expected NPV System Divisions, Massachusetts Institute of Technology.
increases by 85 per cent with flexible mine design, compared to a Cardin, M A, de Neufville, R and Kazakidis, V, 2008. Process to
design without flexibility. The paper illustrates how to incorporate improve expected value of mining operations, Transactions of the
design options (flexibility) into a strategic mine plan in a manner Institutions of Mining and Metallurgy, Mining Technology, 117(2):
that proactively manages inevitable uncertainties. It is hoped this A65-A70.
research will help in justifying more flexible mine designs and de Neufville, R, Scholtes, S and Wang, T, 2005. Real options by
further the sustainability of the industry. spreadsheet: Parking garage case example, Journal of Infrastructure
Systems, 12(2):107-111.
RECOMMENDATIONS Dimitrakopoulos, R and Abdel Sabour, S A, 2007. Evaluating mine plans
under uncertainty: Can the real options make a difference?,
Whilst the model handles a simple case, currently further Resources Policy, 32(3): 116-125.
research and model improvements continue in the following Dimitrakopoulos, R and Ramazan, S, 2004. Uncertainty-based production
scheduling in open pit mining, SME Transactions, 316, 106-112.
areas:
Godoy, M and Dimitrakopoulos, R, 2004. Managing risk and waste
• More detailed modelling which considers multiple process mining in long-term production scheduling, SME Transactions, 316,
options and multi-product options. 43-50.
Leite, A and Dimitrakopoulos, R, 2007. A stochastic optimization model Ramazan, S and Dimitrakopoulos, R, 2004. Traditional and new MIP models
for open pit mine planning: Application and risk analysis at a copper for production scheduling with in-situ grade variability, International
deposit, Transactions of the Institutions of Mining and Metallurgy, Journal of Mining, Reclamation and Environment, 18(2):85-98.
Mining Technology, 116(3):A109-A118. Ramazan, S and Dimitrakopoulos, R, 2007. Stochastic optimisation of
Leite, A and Dimitrakopoulos, R, 2010. Production scheduling under long-term production scheduling for open pit mines with a new
metal uncertainty – Application of stochastic mathematical integer programming formulation, in Orebody Modelling and
programming at an open pit copper mine and comparison to Strategic Mine Planning, second edition (ed: R Dimitrakopoulos),
conventional scheduling, in Advances in Orebody Modelling and pp 385-392 (The Australasian Institute of Mining and Metallurgy:
Strategic Mine Planning I (ed: R Dimitrakopoulos), pp 35-40 (The Melbourne).
Australasian Institute of Mining and Metallurgy: Melbourne). Topal, E, 2008. Evaluation of a mining project using discounted cash
Lima, G A C and Suslick, S B, 2006. Estimating the volatility of mining flow analysis, decision tree analysis, Monte Carlo simulation and
projects considering price and operating cost uncertainties, real options using an example, International Journal of Mining and
Resources Policy, 3:86-94. Mineral Engineering, 1(1):62-76.
Meagher, C, Abdel Sabour, S A and Dimitrakopoulos, R, 2010. Pushback Wang, T and de Neufville, R, 2005. Real Options ‘in’ Projects, paper
design of open pit mines under geological and market uncertainties, presented to Ninth Real Options Annual International Conference
in Advances in Orebody Modelling and Strategic Mine Planning I Paris, France.
(ed: R Dimitrakopoulos), pp 291-298 (The Australasian Institute of Wang, T and de Neufville, R, 2006. Identification of real options ‘in’
Mining and Metallurgy: Melbourne). projects, paper presented to 16th Annual International Symposium of
Morley, C, Snowden, V and Day, D, 1999. Financial impact of the International Council on Systems Engineering (INCOSE),
resource/reserve uncertainty, The Journal of The Southern African Orlando, US.
Institute of Mining and Metallurgy, 6: 293-302.
HOME
Pit Optimisation and Mine Scheduling at the Montes Claros

De Goiás Nickel Laterite Deposit, Brazil
B Y Tomaselli1, E A Silva2, G J de Carvalho Pereira3, J A Hilário4, M C Botelho5 and
P C Rodriguez6
ABSTRACT nickel lateritic deposits: pyrometallurgy and hydrometallurgy

(Davi, Bacon and Osborne, 2004). The choice of the best route
The nickel lateritic deposit of Montes Claros de Goiás (MCG) is located
in the west portion of Goiás state in Brazil. The deposit presents two
will be in accordance to the ore magnesia content and its nickel to
main ore types: limonite and saprolite. The MCG deposit is located on a iron ratio. The degree of inherent homogeneity within a deposit
flat topography and the alteration profile for the deposit presents the will also influence on this decision.
limonite over the saprolite. Limonite presents a lower MgO and Ni grade Typically, a study is conducted considering the best process
and a higher Co grade when compared with the saprolite. A conceptual route to treat the ore presented in the reserve. This approach
analysis was carried out in order to obtain the optimised nickel recovery usually considers only stand alone operations and will leave
through a study comparing two different technologies: High Pressure behind any marginal ore that could be treated by other
Acid Leaching (HPAL) and Ferronickel (FeNi). The methodology used
was to generate a series of pit optimisations and subsequent strategic
technologies. Nowadays some operations that have conducted
mining scheduling, including capital expenditures consideration and cash their projects to stand alone options are considering process the
flow analysis. Considering that the correct selection of the optimum marginal ore with another technology in order to obtain a
ultimate pit is essential for the project profitability, the presented paper maximum recovery of its deposit.
shows that it is suitable in the early definition phase of the project, mainly Considering these two methods for treating a nickel lateritic
during prefeasibility studies. deposit, some characteristics based on operations and projects
should be noted (Brook Hunt, 2007). Ferronickel is a high energy
INTRODUCTION consumption that requires hydroelectric energy for feasibility. It is
only applied for saprolite ores and it has 75 to 90 per cent of
The nickel lateritic deposit of Montes Claros de Goiás (MCG) is nickel recovery and no cobalt recovery. The CAPEX of ferronickel
owned by Votorantim Metais (VM) and is located in the west is in the order of US$28 - 40 thousands per ton of nickel produced.
portion of Goiás state in the city of Montes Claros de Goiás,
HPAL makes large use of reagents and is very sensitive to sulfuric
which is 280 km away from Goiânia, GO, Brazil. Nowadays it has
acid consumption as it usually represents its main reagent
more than 60 000 m of RC drilling holes, bench and pilot
metallurgical tests and is on a prefeasibility stage. Figure 1 shows consumption and may compromise the project feasibility. Usually
the location of the deposit. The total measured and indicated HPAL is only applied to limonite ores with low magnesium
resources estimated are 135.6 Mt at 0.83 per cent Ni and have the content. It consumes less energy and has a significant energy
following distribution for each rock type: limonite are 101.4 Mt at cogeneration on the acid plant. It has more than 90 per cent of
0.72 per cent Ni, 3.6 per cent MgO, and 47.1 per cent Fe2O3, and nickel recovery and 85 to 90 per cent of cobalt recovery and
saprolite are 34.2 Mt at 1.15 per cent Ni, 18.6 per cent MgO and flexible product types. The CAPEX of HPAL is in the order of
21.2 per cent Fe2O3. The choice of the technology route and the US$45 - 60 thousands per ton of nickel produced. If we consider a
final product are directly dependent of the ore characteristics that HPAL stand alone operation at MCG deposit we must consider
will feed the plant, mainly its chemical and mineralogical that the recommended Mg grade is six per cent, as values higher
characteristics. The lateritic deposit composition will be dependent than this will lead to a unprofitable acid consumption and also
on the type of parent rock, the climate where the deposit was scaling problems on the autoclaves. Saprolite usually cannot be
formed and the weathering history. There are two main routes for used on a HPAL operation as it has high magnesia contents,
however, using only limonite and discarding the saprolite is not
recommended as higher Ni grades are on saprolite ore. On the
1. Mining Engineer, Votorantim Metais Ltda, Dept° Geologia/ other hand, if we consider a FeNi stand alone operation we must
Exploração Mineral, Edifício Esplanada, Praça Ramos de Azevedo, consider that limonite is not suitable to be used due to its high iron
206-15° andar – Conjuntos, 1530/1540 Centro, São Paulo 01037-010, content and low nickel grade and a stand alone operation is
Brazil. Email: [email protected]
profitable but also too restricted because of the low mass available
2. Mining Engineer, Geoexplore Consultoria e Serviços, Av Afonso of saprolite when compared to limonite. A complete pit
Pena, 3294 conjunto 206, Cruzeiro, Belo Horizonte MG 30130-009, optimisation study was developed considering both stand alone
Brazil. Email: [email protected] scenarios and also a hybrid operation to find out which one is the
3. Mining Engineer, Votorantim Metais Ltda, Dept° best option. The complete list of the nine scenarios is presented on
Geologia/Exploração Mineral, Edifício Esplanada, Praça Ramos de Table 1.
Azevedo, 206-15° andar – Conjuntos, 1530/1540 Centro, São Paulo
SP CEP 01037-010, Brazil. Email: [email protected] METHODOLOGY
4. Mining Engineer, Geoexplore Consultoria e Serviços, Av Afonso The methodology consists of:
Pena, 3294 conjunto 206, Cruzeiro, Belo Horizonte MG 30130-009,
Brazil. Email: [email protected] • import the geological resource model and apply the dilution
based on a benchmarking on a similar operation;
5. General Manager Mineral Development, Votorantim Metais Ltda,
Dept°, Geologia /Exploração Mineral, Edifício Esplanada, Praça • define the benefit function based on technical and economical
Ramos de Azevedo, 206-15° andar, Conjuntos 1530/1540 Centro, constraints and assign a process route for each rock type;
São Paulo 01037-010, Brazil. • select three notable points among the obtained pit outlines;
Email: [email protected] best Net Present Value (NPV), maximum metal recovered or
6. Mining Engineering Manager, Geoexplore Consultoria e Serviços, largest pit, and a location considered important for project
Av Afonso Pena, 3294 conjunto 206, Cruzeiro, Belo Horizonte MG profitability, 20 years life-of-mine (LOM) operation for
30130-009, Brazil. Email: [email protected] instance; and
B Y TOMASELLI et al
FIG 1 - MCG deposit location.
PIT OPTIMISATION AND MINE SCHEDULING AT THE MONTES CLAROS DE GOIÁS NICKEL LATERITE DEPOSIT, BRAZIL
resource and should not be located on drainages or access

TABLE 1 routes. The bench height selected for this study was 3 m and
Pit optimisation scenarios. the berm width was 4 m resulting on an overall slope angle
of approximately 40°. Figure 2 shows a schematic graphic
Scenario Route Shell representation of the mine considering the predicted destination
1a HPAL Best NPV for each rock type.
1b HPAL 20 years LOM Pit optimisation was performed using Whittle Four-X® to
obtain the benefit of each block (Randall, 1994). A block is
1c HPAL Maximum resource
considered profitable if the income that could be generated is
2a FeNi Best NPV higher than the cost of processing and mining the block and all
2b FeNi 20 years LOM the waste over it (Kennedy, 1990). Because of that, to calculate
2c FeNi Maximum resource
the profitability of a block it is necessary to define all the costs
related and the correlated income that this block could generate.
3a Hybrid Best NPV Process costs included all the supplies, energy, labour, freights
3b Hybrid 20 years LOM and G&A from the crusher till, the client excluding mine
3c Hybrid Maximum resource operation. Mine operation considered all the operations that are
necessary to deliver the ore into the crusher, including all the
auxiliary services. The assumptions that were made are described
• schedule and generate a discounted cash flow (DCF) to below:
compare these scenarios, including the capital Expenditure Costs:
(CAPEX) (Noronha and Gripp, 2001). • FeNi process: flat rate;
• HPAL process: flat rate + acid cost;
GEOLOGICAL MODELLING AND DILUTION
• acid cost: F(Ni, Mg, Al, Fe) × acid price;
The geological modelling for MCG deposit was internally
developed by the Votorantim Metais (VM) geology team. The • mine: loading + transport + auxiliary equipments cost; and
geological interpretation was based on the following chemical • transport cost was considered block-by-block and calculated
criteria: in accordance to the distance of this block to the crusher and
• soil (SO): Ni < 0.3%; Fe ≥ 15%; MgO < 4%; Al2O3 ≥ 10%; adjusted to the bench height of the block.
• limonite (OX): Fe ≥ 22%; MgO < 10%; Fe + Al2O3 ≥ 25%; Income:
• saprolite (SA): Fe < 22%; MgO ≥ 10%; • Ni and Co price: flat long term price,
• bedrock (BD): Ni < 0.6%; MgO > 10; and • FeNi Ni recovery was 90 per cent,
• chalcedony (CD): SiO2 > 70%. • HPAL Ni recovery was 91 per cent, and
The density used on the model was based on an average • HPAL Co recovery was 89 per cent.
obtained by density tests. A diluted model was created based on For HPAL there is also the possibility of upgrading the grades
experience with similar operations. The methodology assumed using a very simple circuit of beneficiation prior to the autoclave
material loss of the upper layer in contact with soil or outcrop, which was considered on this study.
and for the contact zone between saprolite and limonite it was For the determination of the technological cut-off grade, a
considered a dilution of the upper layer with the bottom layer. graphic varying the cut-off Ni grade on the ore was generated.
The dilution was only applied to the blocks that are in the contact The aim was to achieve an average Ni grade of 1.1 per cent on
zone. HPAL and 1.4 per cent on FeNi. The graphs are presented in
Figure 3.
PIT OPTIMISATION Considering the process constraints and the graphs presented,
As it was defined by the geologists, there are two main ore types a cut-off Ni grade of 0.77 per cent was used for limonite with
within the resource: limonite and saprolite. The first one was upgrade. For the saprolite a cut-off Ni grade of 1.0 per cent was
assigned to be processed by HPAL and the second one by FeNi. used. These values were defined as a technological cut-off and
All the other rock types are defined as waste. To be considered a determined by the process area as a target to obtain reasonable
reserve, the block must be classified as measured or indicated results for the plant. All this data was imported into the Whittle
FIG 2 - Schematic representation of the mine operation.
B Y TOMASELLI et al
FIG 3 - Limonite and saprolite grade tonnage curves.
30 k US$/t Ni and the ramp up period: 65 per cent of capacity for

100%
first year. Also considered was the HPAL throughput rate of
90%
1.66 Mt/a, the Mg of six per cent maximum, the CAPEX:
80%
50 k US$/t Ni and the Ramp up period of 25 per cent of capacity
70%
for first year. Finally, the LOM: 20 years of operation (minimum)
60%
was considered. Based on the schedules, a complete discounted
NPV (%)
50%
cash flow analysis (DCF) was carried out in order to generate the
40%
economic drivers for the different scenarios of the project.
30%
Indicators of value (NPV), return (IRR) and investment
20% HPAL FeNi Hybrid efficiency (PVR) were then generated. NPV (Net Present Value)
10%
was obtained for the project considering the opportunity cost of
0%
0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100%
capital for the business and all costs associated. IRR (internal
Mined ore (%) rate or return) is the project rate of growth it is expected to
generate, ie the value on which the NPV of the project is equal to
zero. PVR (Present Value Ratio) was used to evaluate the
FIG 4 - Optimal pit ore tonnage versus NPV considering different efficiency of an investment and can be translated as dollars
technologies. generated (NPV) per dollar invested in terms of maximum
capital exposure. As there are two different technologies with
Four-X® software to run the optimisation. A graphic of all the different capital costs it is important to make use of an
optimised pits was generated using the results obtained by pit economical measure to assist in the best scenario evaluation. The
optimisation from the software. From this graph (Figure 4) three final analysis will lead to three possible scenarios: best HPAL
notable points were distinguished: best NPV, maximum resource, stand alone, best FeNi stand alone and best hybrid operation.
and 20 year LOM. Note that values are regularised to a 100 per Table 3 shows results for all the scenarios. Values for Ni
cent basis for the greatest value obtained. The data for these production, NPV and PVR were regularised to a 100 per cent
points are summarised in Table 2. It is important to note that basis considering the greatest value as 100 per cent.
HPAL grades for this table do not include the upgrade factor and Considering that these are mutually exclusive projects, the
ore mass values are on a 100 per cent basis for the greatest value NPV is the appropriate economical tool to compare them
obtained. (Stermole and Stermole, 2006). Use of IRR or PVR should be
Pit optimisation results show that a stand alone HPAL operation done in an incremental analysis to choose the best alternative. In
leaves behind as waste an important amount of high grade ore that accordance to Tables 2 and 3 the best scenarios are HPAL: 1a,
would probably improve the project NPV. On the other hand, FeNi: 2a and Hybrid: 3a.
using only saprolite ore on a FeNi stand alone operation will mine To check the NPV results an incremental analysis between the
limonite as waste and dump the bigger part of the deposit as best scenarios of each alternative was performed. The incremental
waste. Despite the lower Ni grade of the hybrid scenario when analysis was performed between the hybrid scenarios against each
compared to the FeNi operation, it has the highest nickel recovery one of the stand alone. Results showed that HPAL stand alone is
and the lowest strip ratio compared to stand alone options. the worst option. Also, despite FeNi stand alone having a greater
IRR (Table 3), the incremental analysis showed that the hybrid
MINE SCHEDULING operation presents both positives NPV and PVR. If we consider a
budget constraint this analysis may help to decide the incremental
To complete the comparison between the scenarios, a strategic gain of each scenario against other options. Considering that there
mine scheduling was developed considering the following is no budget constraint, the hybrid scenario presented the best
constraints and assumptions, including the optimisation of Ni performance when compared to the others. Although this will be
grades for the first years and the start-up on second quarter of the considered the base case for the project, the other two scenarios
first year. Also considered were the FeNi throughput rate of will also be studied as a trade-off. The next step was to detail the
0.88 million tonnes per annum (Mt/a), the SiO2/MgO ratio of two mine scheduling for these scenarios considering all the constraints
(maximum), the Fe of 15 per cent maximum, the CAPEX of and also addressing production issues that would make a suitable
PIT OPTIMISATION AND MINE SCHEDULING AT THE MONTES CLAROS DE GOIÁS NICKEL LATERITE DEPOSIT, BRAZIL
TABLE 2
Pit optimisation results for all scenarios (grades in percentage).
Scenario Ore % Ni Co MgO Fe2O3 SiO2 Al2O3 Strip ratio

1a 61.7 1.01 0.066 4.61 47.68 18.23 7.66 2.4
1b 53.4 1.04 0.068 4.50 48.37 17.73 7.52 2.1
1c 64.9 1.00 0.065 4.66 47.37 18.49 7.68 2.8
2a 31.6 1.40 0.030 20.08 21.15 36.11 3.20 2.9
2b 29.3 1.42 0.030 20.15 21.11 36.03 3.19 2.7
2c 34.9 1.37 0.030 19.93 21.33 36.17 3.28 3.5
3a 93.5 1.15 0.055 9.41 39.54 23.69 6.28 1.9
3b 73.4 1.17 0.058 8.86 40.97 22.61 6.35 1.6
3c 100.0 1.13 0.053 9.99 38.27 24.67 6.15 2.1
TABLE 3 global optimum. It is a good project practice that the earlier this
is defined in the project, the less costly it will be to the project in
Mine scheduling economical measures results.
the future. The solution was to assign each rock type to the
Scenario LOM Production NPV IRR PVR process route that it is more suitable. This will ensure that no
years Ni % % % % additional cost will be expended to process that ore and also
1a 22 56 53.8 19.9 20.7
guarantee an optimum reserve. The MCG deposit is very unique,
saprolite has the greatest nickel grades but the limonite mass is
1b 21 56 52.7 20.1 32.1 too high to simply discard. Because of that, a hybrid operation is
1c 23 56 53.6 19.5 32.6 more profitable than just a stand alone. The method is also
2a 22 44 44.1 26.0 49.9 important because sometimes a stand alone operation is not
profitable, but when combined in a hybrid operation the synergy
2b 21 44 43.9 26.4 49.5 between the two could increase profitability. If we consider that
2c 25 44 43.4 25.3 49.0 the capital cost would decrease because of some synergies, the
3a 29 100 100.0 23.8 100.0 difference between the hybrid and the stand alone scenarios
would be greater.
3b 21 100 90.2 23.1 90.1
3c 31 100 83.3 19.0 83.3
REFERENCES
Dalvi, A D, Bacon, W G and Osborne, R C, 2004. The past and the future
and more realistic mine scheduling in terms of operation. That of nickel laterites [online], PDAC 2004 International Convention,
work is not part of this paper. pp 1-27. Available from: <http://www.pdac.ca/pdac/publications/
papers/2004/techprgm-dalvi-bacon.pdf> [Accessed: 29 September
2008].
CONCLUSIONS Kennedy, B A, 1990. Surface Mining, 1206 p (Society for Mining,
Metallurgy and Exploration: Littleton).
It is well known that all lateritic deposits usually present
Noronha, R A and Gripp, A H, 2001. Ultimate pit selection and design, in
different rock types that are suitable to different process routes. Strategic Mine Planning Conference, pp 133-136 (The Australasian
A trade-off study was developed in order to understand it better Institute of Mining and Metallurgy: Melbourne).
and tried to force the treatment of the whole orebody into just Randall, M, 1994. Application of advanced computer-based systems in
one of these process routes. Sometime this approach is valid, but open-pit evaluation, Mining Latin America, pp 77-87 (Institution of
usually it will be necessary to leave behind a significant portion Mining and Metallurgy: London).
of the reserve. It is also very usual to consider only one of the ore Stermole, F J and Stermole, J M, 2006. Economic Evaluation and
types and dismiss the other as waste or marginal ore. This is what Investment Decision Methods, 768 p (Investment Evaluations
normally happens on stand alone operations. This approach is Corporation: Colorado).
also valid, but it is not guaranteed that the reserve will be the
HOME
Technogenic Iron Ore Formations in Siberia as a Secondary

Resource for Production of Ferrous, Rare, Non-Ferrous and
Noble Metals
P A Filippov1, S A Neverov2 and A A Neverov2
ABSTRACT In Kuzbass, the Sheregesh, Tashtagol, Kaz, Abakan

(underground mining), Teisk and Irbinsk (open cut mining)
Five mines and two concentrators supply raw materials to the
metallurgical complex in Siberia, whereas tailings of the iron ore mines are the main suppliers of iron ore to large West Siberian
concentrators are stockpiled at tailing dumps at the mine site. In Siberia, and Novokuznetsk metallurgical complexes. The run-of-mine ore
waste dumps reserve three billion tonnes of wastes and tailings is processed at a crushing and preparation plant to produce a
discharged by mines and ore preparation plants – their volume tends to primary iron ore concentrate of 38 - 44 per cent Fe in grade, then
increase by more than 15 Mt of overburden rocks annually. More than it is fed to the wet magnetic separation circuit with the yield of
3 Mt of the dry and wet magnetic separation tailings are discharged and the marketable product, viz the secondary iron ore concentrate of
stockpiled by mills and ore preparation plants. Open pit dumps keep 62 - 66 per cent Fe in grade at Mundybash and Abagursk ore
600 Mt of overburden rocks. Four underground mines accumulated 50 Mt preparation plants. Tailings are stockpiled at dumps at the mine
of wastes of 0 + 60 mm in size at tailing dumps. Wet magnetic separators site. In total, 373.8 Mt of raw iron ore were produced at Siberian
discharged more than 120 Mt of -2.0 + 0.045 mm slimes. Waste dumps
mines with the yield of 235.8 Mt of primary concentrate.
occupy more than 100 000 m2. Technogenic waste formations contain a
variety of valuable elements and compounds of commercial value. Production of the iron ore concentrates involves the formation of
Utilisation of secondary, technogenic raw materials is rather profitable, as large tailing dumps of 200 Mt, including tailings of preparation
the economic factors of the process are comparable with those for plants and slimes of the wet magnetic separation circuit (Table 1).
processing of mined mineral ores. Qualitative proportions of the valuable In Siberia, six iron ore mines built up 78.7 Mt of wastes from
content in dump wastes were evaluated based on geological exploration crushing and preparation plants in an area of 109 200 m2. The
of aged tailing samples by analytical and geophysical methods. By the large slime dumps are accumulated at ore preparation plants,
assay data, the iron ore tailings contain iron, cobalt, zinc, and other upgrading the primary mine concentrates into concentrates for
valuable metals, and the concentrator slimes have inclusions of precious smelters by employing the wet dressing and dust suppression
metal – gold and silver. The potential of technogenic iron ore reserves at processes. The dumps of Mundybash and Abagursk preparation
mines and ore preparation plants is substantial as a secondary source for
plants amount to 119.2 Mt of the wet separation slimes – 65 per
production of valuable metals in Siberia.
cent of the total wastes are -2.0 + 0.045 mm in size. At present,
the dump materials are classified and sold as construction gravel
INTRODUCTION or utilised as a fill to stow mining cavities under an agreement
Currently in Russia, more than 60 billion tonnes of industrial with the government supervising bodies.
wastes are stockpiled in surface dumps built up by operating
mines, smelters, heat power stations and other works. Most dump TABLE 1
materials are wastes from the mining industry (up to 90 per cent), Waste dumps at iron ore mines and ore preparation plants
smelter slags (two to three per cent), ashes of coal boilerhouses in Siberia.
(one to two per cent) and others (five to seven per cent). More than
75 per cent of mining wastes fall on overburden rocks at open pits Process Name Volume Area Valuable
(Mt) (m2) elements
and up to 15 per cent are tailings, discharged at ore, coal, or rock
preparation plants. Approximately three billion tonnes of rocks Mines
produced by opencast mining are processed annually, with more Underground Sheregesh 23.32 50 900 Fe, Mn, Zn
than 500 Mt of overburden rocks being mined at iron ore open mining, dry
Kaz 7.2 6300 Cu, Zn
pits. In total, about 800 Mt of tailings are stockpiled at dumps of magnetic
iron ore mines in the Siberian region. Mine dumps are increased separation Tashtagol 0.3 200 Fe
by more than five million cubic metres of overburden rocks Abakan 20.1 28 000 Co, Cu, Ni
annually – the crushing and concentrating factories discharge Open cut mining, Teisk 21.2 19 100 Co, Zn
above 2 Mt of tailings. This intensive mining activity results in a dry magnetic
heavy ecological impact on the local environment. More than 40 Irbinsk 6.7 4700 Co, Cu, Zn
separation
per cent of disturbed lands, 30 per cent of harmful atmospheric Ore preparation plants
emissions and up to 20 per cent of sewage water release into
natural water basins in Russia are because of mining operations. Wet magnetic Mundybash 30.1 60 100 Au, Ag, Zn
Of special note is the ecological load on the Kuznetsk Basin, one separation
Abagursk 89.1 385 200 ZN, Cr, Co
of the leading metal, coal, ferrous and non-ferrous producers in Total: 198.1 554 500 -
the Russian Federation.
The Institute of Mining of the Siberian Branch of the Russian

1. Senior Researcher, Institute of Mining, SB RAS, 54 Krasny Academy of Sciences performed the survey studies aimed at
Prospect, 630091 Novosibirsk, Russia. Email: [email protected] assessing the composition and properties of mine dump materials.
It was concluded by the exploration data that most dump materials
2. Researcher, Institute of Mining, SB RAS, 54 Krasny Prospect, can be reprocessed and utilised as a secondary raw material
630091 Novosibirsk, Russia. Email: [email protected] (Filippov, Uskov and Freidin, 2003; Gaidin et al, 2003, Freidin,
3. Technogenic formation is a dump formed by many years of dumping Filippov and Gaidin, 2002). Moreover, utilisation of technogenic3
industrial wastes that could potentially have a commercial value to raw materials can be economically profitable and comparable with
produce new products by newly developed methods. the processing of raw ore mined at geogenic mineral deposits
P A FILIPPOV, S A NEVEROV and A A NEVEROV
(Doroguntseva, 2006). The commercial evaluation of technogenic and ore preparation plant on a 15 - 20° slope, contains 1.6 Mt of
formations comprises a large-scale comprehensive exploration that ore wastes of variable iron content (Figure 1). The dump
includes both laboratory and in situ tests, as well as the stretches for 300 m, the maximum width of its horizontal section
development of standards for raw mineral materials, with reaches 40 m and the tailing bed is within 25 - 30 m. The iron
consideration for the summarised economic efficiency of the content in the Feofanovsky dump material is summarised in
utilisation of technogenic wastes. The ecological factor, involving Table 2. It is explicit that iron content in the upper layer of
preventing the release and discharge of hazardous components and consolidated tailings is higher by 12 per cent as compared to an
compounds into water basins and the atmosphere, should be average value, and tends to increase with the decrease in fraction
assessed along with the economic value of technogenic mineral size to 21.4 per cent at average iron content of 15.2 per cent. The
resources. Of particular economic and social importance is the fine fraction (zero to 15 mm) makes up one third of the total
restoration of disturbed lands in agricultural and recreation areas.
The magnetic properties of the aged iron ore tailings become
substantially weaker. The relatively high content of sulfide TABLE 2
minerals contributes to intensive oxidation processes followed by Iron distribution in fraction sizes.
liberation of gold, cobalt, nickel and other metals from the crystal Size fraction (mm) Ratio (%) Fe content (%)
lattice. These oxidation processes entail formation of ecologically
hazardous compounds of sulfur, arsenic and other elements. In Larger than 50 22.4 13.3
view of the above, processing of technogenic dump resources is 40 - 50 8.1 12.4
actualised through economic, ecological and social aspects. 20 - 40 20.5 13.2
The iron ore dumps at the Sheregesh Mine are of special
15 - 20 17.6 15.0
commercial value. Their reserves are above 20 Mt of tailings
with more than 20 per cent Fe in grade (Filippov, 2008). The 10 - 15 11.3 16.8
scientific and in situ investigations into production of the iron 5 - 10 16.3 19.2
ore concentrate from tailings of the Sheregesh Mine were
Less than five 3.7 21.4
performed at the Feofanovsky dump, favourably located in the
production mine site. The waste dump, 3 km from the crushing Total 100.0 Average 15.2
Fe content, %
FIG 1 - Isolines of iron content in mill tailings, Feofanovsky waste dump.
TECHNOGENIC IRON ORE FORMATIONS IN SIBERIA AS A SECONDARY RESOURCE
volume with iron content Fe = 19.1 per cent. The large size workload level no doubt hampers the processing of aged tailings.
fraction (>50 mm) constitutes one fifth of the total volume and is The problem can be solved by employing a mobile modular
relatively poor in iron – large-size lumps should be separated, complex, designed to produce an iron ore concentrate from old
crushed and processed by dry magnetic separation. consolidated tailings (Figure 3). The mobile complex can be
mounted at the waste dump site. It consists of a receiving bin,
CASE STUDY two one-drum magnetic separators, screening circuit for material
classification into size fractions and a belt-conveyor line for
The newly developed process flow sheet (Figure 2) was employed transportation of concentrated products. The primary material
in the full-scale tests for Feofanovsky-aged dump material at the is unloaded by excavator, bulldozer or scraper facilities into a
Sheregesh crushing-preparation plant. The tests were scheduled on receiving bin. The choice of loaders and trucks depends on the
Sunday, when the factory exclusively processed dump material capacity of a mobile concentrating complex and its operation mode.
delivered by trucks. The dump material was unloaded first into a
The construction of the modular concentrating complex is
special bin at the factory, then fed into bunker No 36 in the second
launched at the Sheregesh Mine to process 12 divided by 14 per
technological line. The material was fed to separator Number 52
cent Fe tailings of Kumnagash dump, whose reserves are 9.1 Mt.
and then with no processing at the separator, it was transported by
It is planned to produce an intermediate product (Fe = 35 per cent)
conveyor lines Number 60, 49, 51, 53 and 55 to crusher Number
for further treatment at Sheredesh and Mundybash ore preparation
15 in the first technological line. The material ground down to 16
plants, with the yield of the finished concentrate (Fe = 42 per cent
to zero millimetres was fed to screen Number 19, the oversize
or Fe = 62 per cent). The investment cost of the project is
fraction was re-crushed in fine crusher Number 23 and the
US$820 000 including construction, assembly works and purchase
undersize was moved to conveyor Number 25. The re-crushed and
of mineral dressing equipment. The waste- concentrating complex
undersized products were combined on conveyor Number 25 and
consisting of a concentrator module, diesel loader, five trucks and
fed to screen Number 27, which is equipped with three levels of
a bulldozer, is served by a 25-man team. Its project production
different-mesh sieves (20 × 20 mm upper, 18 × 18 mm middle and
capacity is 250 000 t of intermediate product and 130 000 t of the
13 × 13 mm lower sieves). After screening the oversize product of
primary concentrate at the annual planned processing rate of one
+16 mm, it was conveyed by Number 51, 53 and 55 to crusher
billion tonnes of wastes at Sheregesh preparation plant. The total
Number 15, while the undersized product was conveyed by
design production cost of the concentrate is US$8.7 per one tonne,
Number 29 to magnetic separators Number 45 and 47. The
including US$3.2 per one tonne of the intermediate product at an
prepared feed was subjected to the dry magnetic separation at
ore preparation plant, and US$4.4 per one tonne of the initial feed
separators at controlled operating modes. The final product
at the concentrating complex. The project payback period is one
(primary concentrate) was conveyed by Number 61, 62 and 66 to
year for Kumnagash tailings. The process also involves the
the finished product storage area. The tailings were conveyed by
production of a marketable by-product, namely, classified
Number 57, 59, 63, 65, 67, 71, 73 and 75 to screens Numbered 65
construction gravel. Furthermore, the process provides higher
and 69 for classification to yield the finished construction material.
efficiency, more complete recovery of valuable components and
Thus, the primary iron ore concentrate of the required grade was
less ecological impact, thanks to the optimal operating modes of
produced in the full-scale tests of four bulk samples of
magnetic separation of iron ore tailings, heterogeneous in
Feofanovsky-aged dump material with the use of magnetic
granulometric composition and content of valuable components.
separators Number 45 and 47 (Table 3).
The exploration survey of Siberian mines revealed the adequacy in
By the test evidence, the intermediate product, viz a primary their qualitative characteristics and proved the satisfactory
iron ore concentrate of up to 36.9 per cent Fetotal in grade, was reliability in interpretation of the Fe recovery data obtained for
produced under industrial conditions from old consolidated one waste dump to others.
tailings at the average concentrate yield of 20.8 per cent and The Mundybash slime dump accumulates 30 Mt of iron
recovery of 48.3 per cent for separator Number 45. Concentrates bearing slimes of wet concentration and the fine-grain material
of 36.8 per cent Fetotal in grade with the average concentrate yield sized -2.0 + 0.045 mm makes up 65 per cent of the total mass. In
of 26.8 per cent and recovery of 61.4 per cent were produced at evidence derived from the comprehensive analysis, Zn, Cu, Co
separator Number 47. The evaluated cost of one tonne of 33.4 and Mn elements prevail in the slimes. Of great value is the noble
divided by 36.9 per cent Fe concentrate produced from current metal (gold, silver) content in slimes – these metals are distributed
tailings of the preparation plant was US$3.2. The actual cost of irregularly around the dump area and their content in core
one tonne of the primary concentrate of 41 per cent Fe in grade specimens varies – Au = 0.1 - 1.2 g/t (weighted average value is
produced from a crude ore (26.4 per cent Fe) of the Sheregesh 0.616) and Ag = 0.5 - 7.0 g/t. The centrifugal separation provides
underground mine was US$19.7. It is important to point out that the maximum gold recovery of 93 per cent at the gold
the cost of one tonne of 36.8 per cent Fe primary concentrate concentration ratio of 8:23 versus the initial gold content in a
produced from the ore of Teisky open pit is US$16.1. In Siberia, sample under laboratory conditions. The process is applicable to
the production capacity of crushing – preparation plants is recover silver, sulfide concentrates and finely-dispersed
utilised by 80 - 95 per cent at all operating mines. Such a construction materials along with the base gold.
TABLE 3
Assay data on the concentrate and tailings samples at Sheregesh preparation plant.
Separator Test Feed Concentrate Tailings

number number
Fetotal (%) Femag (%) Fetotal (%) Femag (%) Yield (%) Recovery (%) Fetotal (%) Femag (%)
1 15.9 8.9 36.4 30.3 22.9 52.4 9.8 3.9
Number 45 2 15.8 9.2 36.9 30.0 19.2 44.8 10.8 2.9
Average 15.8 9.1 36.7 30.2 20.8 48.3 10.3 3.4
1 15.3 9.3 36.8 31.3 22.7 54.6 9.0 1.8
Number 47 2 17.3 10.7 33.4 27.8 31.8 61.4 9.8 2.3
Average 16.3 10.0 35.1 29.6 26.8 57.7 9.4 2.1
P A FILIPPOV, S A NEVEROV and A A NEVEROV
FIG 2 - The in situ test flow sheet for processing of current tailings of Sheregesh crushing-preparation plant.
TECHNOGENIC IRON ORE FORMATIONS IN SIBERIA AS A SECONDARY RESOURCE
Wastes CONCLUSIONS
1 In Siberia, about 200 Mt of iron ore dressing process wastes are
stockpiled – the wastes currently occupy an area of more than
100 000 square metres. It has been found that technogenic
2 formations of aged mine wastes and tailings contain potentially
valuable ferrous, non-ferrous and noble metals (Fe, Co, Zn, Au
and Ag). The reasonability of development and implementation
4 of innovative projects for the processing of technogenic waste is
3 apparent, as modern mineral-dressing processes allow the
efficient processing of available dump materials to produce
marketable products: iron ore concentrate, base gold, sulfide
concentrates, and construction materials.
The planned concentrator for recovery of gold and sulfide
5 concentrates from current and aged tailings is based on
the exploration studies carried out at the Mundybash ore
preparation plant.
6 REFERENCES
Doroguntseva, N P, 2006. Development of comprehensive utilisation of
natural and secondary mineral resources, based on diversification of
activity of mining enterprises (in Russian), PhD thesis, Novosibirsk.
Filippov, P A, 2008. Potentials of technogenic formations in mines of
5 West Siberia, Journal of Mining Science, 4:386-390.
2 Filippov, P A, Uskov, V A and Freidin, A M, 2003. Perspectives for
processing of technogenic raw materials at the Sheregesh Mine (in
Russian), in Science Intensive Processes for Mining and Beneficiation
of Mineral Resources, pp 164-166 (IGD SO RAN: Novosibirsk).
Freidin, A M, Filippov, P A and Gaidin, A P, 2002. Process for treatment
of iron ore tailings, Russian Federation Patent No 2190027.
Construction material Concentrate (Fe=35%) published in Bulletin of Inventions, No 27.
Gaidin, A P, Filippov, P A, Komissarov, A V and Onofreichuk, V Ya,
2003. Treatment of old consolidated tailing dump materials in the dry
1 - Receiving bin; 2 - Doser ; 3 - Screen ; magnetic separation cycle at a concentrating factory of OAO
4 - Magnetic separator; 5 - Conveyor ; 6 - Discharge bin . ‘Sheregesh Mining Board’ as a factor improving the regional
ecological situation and economic efficiency of iron ore processing,
in Proceedings Fourth Congress of Mineral Processing Experts of
FIG 3 - Modular complex for processing of tailings. CIS Countries, pp 192-193 (MISiS: Moscow).
HOME
Advancing Strategic Mine Planning – A Software Perspective

C Tonkin1
ABSTRACT and better able to handle increasingly complex problems. We

have come a long way along the optimisation pathway – for
Strategic mine planning for open pit mines is a mature process, in which
sophisticated software tools are used to optimise everything, from pit example, expanding pit optimisation to cut-off and blend
design, to pushbacks, to scheduling. Since inputs into the planning optimisation – and we have a good understanding of the business
process are complex and uncertain, current practices call for breaking the processes related to strategic mine planning.
task into component parts and addressing them sequentially. Research This paper provides the context surrounding strategic mine
suggests that this paradigm may soon change, with the expanded planning and analyses current techniques that are applied to the
application of such concepts as simultaneous planning. In addition, problem space. Detailing the history of advances in the area to
techniques for managing the unknown inputs of metal content and future date, the paper describes innovative solutions and their
financial conditions – notably stochastic simulation and real options –
relationship to the complexity of the mine planning problem.
promise to leverage uncertainty in order to deliver increased mine value.
While to date techniques focus primarily on open pit mines, they lend Finally, a number of technologies are discussed for advancing
themselves to a logical extension into the underground mining environment strategic mine planning in the future.
as well.
The complexity of strategic mine planning
INTRODUCTION Strategic mine planning is concerned with the overall direction
Strategic mine planning is the art of designing a business strategy of the mining company. Based on a clear understanding of the
that will deliver maximum business value in extracting an market and relevant economic factors (price, cost, and discount
available resource, through the optimal allocation of time, rate assumptions), it seeks to optimise the allocation of scarce
capital, technologies and people. For more than two decades, the capital resources by aligning physical parameters such as the
global mining industry has used increasingly sophisticated grade and tonnes or metal produced with the requirements of the
mathematical methods and techniques implemented in software business. Business requirements might include positive cash
packages to optimise this planning process. However, strategic flows throughout the mine life, a hurdle rate for investment that
mine planning depends on a complex, interdependent, and highly must be satisfied, a minimum amount of metal to be delivered at
uncertain system of inputs. Some uncertainties relate to the a given time, or a minimum grade/quality target for the final
resource itself, such as grade and tonnage of material or metal in product.
the ground, and geotechnical stability of the resulting excavation, Some key questions related to the mining operation include:
and others including uncertainty in the operating conditions that • What capital configuration and associated cost model will
will prevail during production and capital required. We also live maximise the return on investment?
with the fact that, although a significant upfront investment is
required to start a mining project, the resource itself will be • What are the implications of different financial and slope
extracted over a period of five, ten, 20, or 30 years, or even scenarios, and should the final design be adjusted to suit?
longer. The return on investment will therefore be delayed into • To what parameters is the deposit most sensitive, and is
future and uncertain times. further investment needed before commencing the mining
A key challenge presently facing the mining industry lies in operation?
optimising the mine as an integrated process. Ideally, all relevant Strategic mine planning addresses these questions and more,
factors should be considered and addressed simultaneously, from enabling the mine planner to make decisions about the final
the perspective of understanding how they influence the strategic strategy and ultimate mine configuration based on sound
objective for the business – typically maximum net present value analysis.
(NPV) for the mine. What type and grade of material should be
Pit optimisation alone does not provide sufficient information
extracted? When is the best time to extract this material? Where
on the mine plan to enable sound decision-making. The final
is the best place to begin mining? How should the mine be
mine strategy decision goes well beyond the size and shape of
staged, such that the strategy can be readily modified if
the final pit by looking at the mine as a whole – pit, schedule,
conditions change? Can the various elements in the deposit be
sequence, staging, plant, port, ore waste discrimination,
blended to create an optimal product mix for the market? What
blending, and pushback selection. By identifying a range of
risk reduction strategies should be followed – mining a smaller
viable options, investment decisions can be made according to
extraction shape first to minimise the impact of future price
the requirements of the individual situation. To do this
uncertainty, carrying hedge books, improving data inputs through
effectively, a range of advanced planning tools and rapid
increased drilling density? These questions and more relate to
evaluation of different scenarios are required.
optimising the mining process, and a similar number of questions
can be raised for the optimisation of the downstream processing Many different planning options are available, depending on
and logistics for the mine product. the size of the final pit, as shown by Whittle (1997). Figure 1
depicts several possible strategies, including maximum resource
Because future business and market conditions can only be
point, maximum NPV point, and a range of less risky designs
estimated, planning must be based on the best possible
requiring a smaller capital outlay. The shaded area in the diagram
predictions related to cost, price, and other key factors. Today’s
represents all financially viable and technically feasible mine
strategic mine planning practices do a good job, with excellent
designs. Looking at the far right side of the graph, the decision to
contributions on both the technological and business side.
extract the maximum economically viable reserves represents an
Current software tools make the planning process faster, easier,
aggressive approach. The risk-neutral mine design at the top of
the curve would exploit the resource for the purpose of returning
1. Technical Product Manager – Scheduling and Optimisation, maximum NPV to investors, sacrificing total ore tonnes. Farther
Gemcom Software International Inc. left is a range of risk-averse designs that reduce capital investment
C TONKIN
on investment, reserve targets, grade consistency, geotechnical

uncertainty, capital investment, commodity prices, stockpiles –
all these and more make planning optimisation an ongoing
challenge.
Current strategic mine planning practices lay the

foundation for their advancement
The standard approach to strategic mine planning is to break the
overall task into discrete, manageable pieces. For example, mine
planners look at different slope scenarios, using arc relationships
between blocks in the model, use costs and revenues to
determine ore waste discrimination and block values, and employ
Lerchs Grossman (Lerchs and Grossman, 1965; Whittle, 1988)
pit optimisation routines to create a range of options for mining
the orebody. Next comes strategic scheduling where planners
typically follow a step-wise process of pushback selection,
FIG 1 - Making resource allocation decisions.
staging the mine to expose the ore, balance stripping requirements,
and improve return on investment. They seek to maximise NPV
and mine life but still return a positive NPV. It is important to at the marginal cut-off, then investigate further value by looking
note that any of these alternatives might be considered optimal, at cut-off optimisation and blending strategies.
depending on a company’s resource allocation strategy. In this The development timeline of Gemcom Whittle, a long-time
example, the information needed to make these critical resource leader in strategic mine planning software, provides insight into
allocation decisions, along with drill-down capabilities that the industry trends that have led to the practices and techniques
provide more detail on the capital and operating cost, price, and that are widely used today. Whittle brought the first commercial
recovery inputs for any given option, was generated in Gemcom three-dimensional pit optimisation package to market in 1985 –
Whittle™ software. Having this level of detail visible, accessible, 20 years after publication of the Lerchs Grossman pit
and auditable allows the mine planner to rapidly analyse the optimisation theory (Lerchs and Grossman, 1965) – allowing its
different options and make the best choice for a given case. users to quickly and accurately calculate the optimal pit for a
The extraction strategy that is selected will have a direct impact given resource. This hitherto unavailable ability to calculate the
on the cash flow of the mine. In Figure 2, for example, the optimal pit through an automated process provided mine
decision has been made to maximise NPV within the constraints planners with both a mathematically optimal solution and an
of the final pit design. The graph shows the effect of different enormous time saving compared to manual processes. Using
sequencing approaches, with the two lines representing two Whittle for pit optimisation, the mine planner now had a process
different extraction strategies with exactly the same final pit, that provided the maximum value solution for the final pit. This
resources, ore, and waste tonnes. Depending on the extraction enabled more time to be spent on accurately modelling costs,
strategy, the project will deliver vastly different values. Strategic developing alternative scenarios and, for the first time, analysing
mine planning is a complex process, and the many variables the sensitivity of the deposit to changes in input parameters, most
preclude a simple mathematical solution. Mine life, NPV, return importantly commodity prices. (Whittle, 1988)
FIG 2 - Identifying the best strategy for any given decision.
ADVANCING STRATEGIC MINE PLANNING – A SOFTWARE PERSPECTIVE
The following year, Whittle added the fourth dimension, time,

which enabled discounted cash flow to be calculated. This
enhancement enabled the mine planner to make decisions that
would improve the NPV and return on investment for a deposit. In
1995, Whittle incorporated Lane’s cut-off optimisation theory
(Lane, 1988; Whittle and Wharton, 1995), according to which cut-
offs are varied over time to maximise the NPV for a mining
sequence. By ‘productising’ published theories through commer-
cially available software, Whittle thus created a core set of
building blocks for strategic mine planning – pit optimisation,
economic analysis, and cut-off optimisation – that delivered the
capability to determine optimal solutions for a deposit and to
examine different mining scenarios and sensitivities. The years on
either side of the millennium – the late 1990s and early 2000s –
marked a new stage in the industry, with innovative solutions join-
ing the basic building blocks to deliver the benefits of Whittle to a
wider range of projects and unlock even greater value with
additional optimisation routines. Whittle development continued
to reflect the needs of the industry, beginning with the introduction
of multi-element deposit capability. This feature clearly showed
revenue contributions from more than one element, giving poly-
metallic deposits the same benefits enjoyed by single-element
deposits, that is, pit optimisation, economic analysis, and cut-off
optimisation. To enable mine planners to achieve maximum NPV,
Whittle developers undertook a unique research initiative resulting
in a proprietary schedule optimisation technique through
its Milawa Algorithm module (Wharton, 2000). The Milawa
Algorithm, results output example shown in Figure 3, provides
automated schedule optimisation routines that maximise the NPV FIG 4 - Whittle’s Navigation Tree organises projects into an
for a given final pit. With all the complexities involved in strategic easy-to-use structure.
mine planning and the utilisation, manipulation, and management
of large amounts of data and end results, software ease of use, the enables users to generate the information they need and even test
logical organisation of Whittle projects is shown in Figure 4, hundreds of different optimisation scenarios quickly, in order to
became the next area addressed by Whittle. find the one that will deliver the best results. Today, Whittle is
Considerable research went into developing the unique, widely known for its ease of use. Following this, Whittle’s
logically organised project environment of the software, which Blending, Stockpiles and PushBack Chooser modules came to
FIG 3 - Milawa schedule to maximise net present value.
C TONKIN
market in rapid succession. These modules enabled more expanded the methods available for calculating the value of a
detailed optimisation and analysis of the mining process faced by block. Companies could now use predefined block values or,
companies. The Blending module delivered the ability to alternatively, custom calculations to value blocks. This capability
optimise blended final products or to blend the input to the provided virtually unlimited ways in which the software could be
processing plant (Wharton, 2004). The PushBack Chooser applied, and the level of complexity it could model and optimise.
module automatically selected the optimal NPV set of pitshells Value expressions unlocked the potential for pit optimisation to
for any given final pit (Figure 5). The Stockpiles module work on dynamic block values.
delivered the capability to model the use of stockpiles in various Next came high-performance binary file support for growing
ways, a critical part of strategic mine planning for some data sets, plus a series of new technologies designed to expand
operations. Now, the industry had automatic scheduling routines memory and limits, streamline processes, support block
to optimise the sequence for NPV, not just report it, and to calculations, maximise NPV, and more. The drive to deliver
optimise for a range of commodities, final products, and optimal solutions, combined with practical mining scenarios, has
pushback and stockpile strategies. The software was then culminated in developments such as the NPV Practical PushBacks
expanded to work with data from different systems, further module. This module automatically creates a set of pushbacks
improving the user experience. Working from the building blocks that address maximum NPV and minimum mining width
in the previous stage, the millennium years saw the creation of a requirements, providing fast and accurate methods for optimising
total solution for open pit mine planning. NPV with more practical constraints. In summary, Whittle’s
advancements have responded to the industry’s demand for more
From 2003 to the present, Whittle and the mining industry have
sophisticated tools that can be applied to an ever wider range of
stretched the limits of strategic mine planning to solve ever more
strategic mine planning challenges and project types. The focus
complex problems. A raft of features have added new value to the
has remained on providing mathematically optimal solutions to
market, including Whittle’s Multi-Mine module to optimise not
meet the real-world challenges faced by mining companies.
just a pit, but an entire mine operation consisting of multiple pits
and all related products and processes (Hall, 2004). Optimising
multiple mines simultaneously helped exploit synergies between Moving into the future
mines, making it possible to maximise NPV, balance stripping Looking at the Whittle example, it is clear that strategic mine
requirements, and achieve grade targets. To tailor the software to planning theory has been on a trajectory of virtually exponential
individual company needs, the Value Expressions module development, as shown in Figure 6.
This trajectory is good news for the mining industry, because it
means we should be able to look forward to significant advance-
ments in the near future. As we have seen, today’s strategic mine
planners generally break problems down into smaller pieces,
optimise discrete components, and solve individual parts. By
doing so, they are able to exploit mineral deposits quite effectively.
But as our ability to model more realistic scenarios grows, and as
our input models become increasingly complex and inter-
dependent, this sequential approach will fall short when it comes
to optimising mine value. What are some of the current research
and possible future technologies that may take the mining industry
to the next level? One exciting research area looks at simultaneous
planning of building blocks, making it possible to take a more
holistic approach to the overall planning process. The goal is to
integrate each building block and optimisation process to solve the
mine planning process simultaneously.
Research in this area continues, using frameworks such as new
Mixed Integer Linear Programming (MILP) models (Ramazan,
2007; Stone et al, 2007) and global optimisation techniques
(Whittle, 2009). Limited commercial solutions for simultaneous
FIG 5 - Create net present value practical pushbacks in a planning already exist: Some Gemcom clients in the coal
single step. industry have been using a commercial MILP solution for a
FIG 6 - Timeline for Gemcom Whittle.
ADVANCING STRATEGIC MINE PLANNING – A SOFTWARE PERSPECTIVE
examine a solution space of many possible outcomes based on

multiple realisations of the orebody, for example, real options can
be used to help determine the best candidate. (Groeneveld, Topal
and Leenders, 2009; Abdel Sabour and Dimitrakopolous, 2010, in
this volume). While price is the biggest influence on profit, it is
not constant. How can we model future prices and discount rates?
Current practice reduces the price into the future, in order to
mitigate some of this uncertainty. An alternative is to use option
techniques to model the effect of changing the mine plan in
response to future, uncertain price movements.
Moving underground
Strategic mine planning has focused primarily on open pit
environments so far, but there is a logical extension to
underground mining as well. In Whittle, for example, there are
already techniques to help determine the transition point between
surface and underground mining. Providing information to the pit
optimiser that an underground opportunity exists, influences the
decision-making regarding final pit size. Moving to the
FIG 7 - Simultaneous planning. underground mine itself, the opportunity exists for more planning
tools to optimise various parts of the mine planning process. The
strategic mine planning process for underground is similar to that
number of years. While current technologies go some way for open pit. It considers the size and shape of the final resource,
toward solving mine planning processes simultaneously, they all the ore waste discrimination point, the best way to stage the mine
take pit optimisation as an input to the process and involve other and extract the resource, and finally the optimisation of the
building blocks to varying degrees. Moving forward, solutions mining sequence itself. As with open pit, individual parts of the
will need to deal with schedules, stockpiles, ore waste underground process – for example, stope sizing and decline
discrimination, and blending with even greater power and placement – can be optimised separately. Currently, this planning
sophistication. Simultaneous planning through technologies such is typically done manually, although Gemcom PCBC™ provides
as MILP solutions and global optimisation are important optimisers for block caving operations, including cave footprint
developments, which the industry can hopefully look forward to finders and a schedule optimiser. Current research in the industry
exploiting in the coming years. is looking at ways to optimise each step in the mine planning
process (Brazil et al, 2009), filling in the gaps with manual
techniques where the tools do not yet exist. This is similar to the
Leveraging uncertainty building blocks stage of open pit strategic mine planning.
Other research focuses on the uncertainty and risk inherent in Research projects are now examining ways to apply some open
strategic mine planning, for example in geological, geotechnical, pit mine planning advances to underground mining: MILP and
operational, and price models. Uncertainty and risk can be other formulations can be used to optimise schedules for
reduced by improving the inputs to the models – more drilling underground (Elkington, Durham and Myers, 2009) and metal
would provide a better assessment in the geological model, for uncertainty techniques also have potential applicability (Grieco
example - while additional drilling and analysis would enhance and Dimitrakopolous, 2007). Since underground mining can
the accuracy of geotechnical predictions. Reducing uncertainty leverage the advancements made for open pit, it seems reasonable
in this way is costly and reaches the point of diminishing return; to expect that the timeline should be significantly condensed.
at best, it can reduce but not eliminate the risk.
Still further research suggests the use of techniques, built on CONCLUSION
foundations of sensitivity and scenario analysis, for working with It has taken 30 years to evolve from manual plans to our current
uncertainty. These types of techniques can help to evaluate the level of sophistication in open pit strategic mine planning. Many
consequences of uncertain inputs. Grade and metal uncertainty excellent tools and techniques are now available, but they tend to
may also be evaluated using conditionally simulated models, be sequential and siloed. In the near future, it is likely that the
quantifying the risk associated with a given mine plan or industry will start leveraging sophisticated methods to incorporate
choosing mine plans that operate well under uncertainty; hybrid uncertainty into decision-making processes, along with simul-
pit techniques and risk-based optimisation frameworks are taneous planning technologies. Underground strategic mine
examples of this approach (Whittle, 2007; Dimitrakopolous, planning, while currently not as advanced as open pit, is expected
Martinez and Ramazan, 2007). Current research into simulated to move along the same path at a more advanced rate.
annealing and stochastic integer programming methods for
production scheduling (Godoy and Dimitrakopolous, 2007; Regardless of future direction, however, one basic tenet will
Ramazan and Dimitrakopolous, 2007) also offers the ability to not change: Within the strategic mine planning framework,
leverage uncertainty in mine planning. Using uncertainty to drive flexibility must be ensured to determine different mine plans
mine planning decisions represents a leap forward from creating based on business objectives, as not every company will opt for
mine plans based solely on deterministic inputs. maximum NPV. Robust mine plans, managerial flexibility, and
innovative software solutions make this a truly exciting time to
What about uncertainty in operational and financial models? be in the mining industry.
Research in this area takes its cue from the financial community in
the form of the ‘real options’ evaluation technique, which
recognises that managerial decisions can and will be taken in the REFERENCES
future to react to changing circumstances. This method essentially Abdel Sabour, S A and Dimitrakopolous, R, 2010. Accounting for joint ore
quantifies the value of management flexibility in a world of supply, metal price and exchange rate uncertainties in mine design, in
change and uncertainty. Mine plans that permit different options Advances in Orebody Modelling and Strategic Mine Planning I (ed: R
may have higher value than those that do not; in other words, the Dimitrakopolous), pp 29-34 (The Australasian Institute of Mining and
option to take certain actions has a value in and of itself. When we Metallurgy: Melbourne).
C TONKIN
Brazil, M, Grossman, P A, Lee, D H, Rubenstein, J H, Thomas, D A and Ramazan, S, 2007. Large-scale production scheduling with the
Wormald, N C, 2009. Access optimisation tools in underground mine fundamental tree algorithm, in Orebody Modelling and Strategic
design, in Proceedings Orebody Modelling and Strategic Mine Mine Planning, second edition (ed: R Dimitrakopolous), pp 121-127
Planning (ed: R Dimitrakopolous), pp 237-241 (The Australasian (The Australasian Institute of Mining and Metallurgy: Melbourne).
Ramazan, S and Dimitrakopolous, R, 2007. Stochastic optimisation of
Dimitrakopolous, R, Martinez, L and Ramazan, S, 2007. Optimising open long-term production scheduling for open pit mines with a new
pit design with simulated orebodies and Whittle Four-X – A integer programming function, in Orebody Modelling and Strategic
maximum upside/minimum downside approach, in Orebody Mine Planning, second edition (ed: R Dimitrakopolous), pp 385-391
Modelling and Strategic Mine Planning, second edition (ed: R (The Australasian Institute of Mining and Metallurgy: Melbourne).
Dimitrakopolous), pp 201-206 (The Australasian Institute of Mining
Stone, P, Froyland, G, Menabde, M, Law, B, Pasyar, R and Monkhouse,
P, 2007. Blasor – Blended iron ore mine planning optimisation at
Elkington, T, Durham, R and Myers, P, 2009. Optimising value for an Yandi, Western Australia, in Orebody Modelling and Strategic Mine
underground project configuration, in Proceedings Orebody Planning, second edition (ed: R Dimitrakopolous), pp 133-136 (The
Modelling and Strategic Mine Planning (ed: R Dimitrakopolous), Australasian Institute of Mining and Metallurgy: Melbourne).
Wharton, C, 2000. Add value to your mine through improved long term
Melbourne).
scheduling, in Proceedings Whittle North American Strategic Mine
Godoy, M and Dimitrakopolous, R, 2007. A multi-stage approach to Planning Conference, Breckenridge.
profitable risk management for strategic planning in open pit mines,
Wharton, C, 2004. The use of extractive blending optimisation for
improved profitability, in Proceedings Orebody Modelling and
(ed: R Dimitrakopolous), pp 337-344 (The Australasian Institute of
Strategic Mine Planning (ed: R Dimitrakopolous), pp 69-76 (The
Grieco, N and Dimitrakopolous, R, 2007. Grade uncertainty in stope
Whittle, D, 1997. The relationship between economic design objectives
design – Improving the optimisation process, in Orebody
and reserve estimates, in Proceedings Resource to Reserve Inputs
Modelling and Strategic Mine Planning, second edition (ed: R
Seminar, pp 27-29 (The Australasian Institute of Mining and
Metallurgy: Melbourne).
Whittle, D and Bozorgebrahimi, A, 2007. Hybrid pits – Linking
Groeneveld, B, Topal, E and Leenders, B, 2009. A new methodology for
conditional simulation and Lerchs-Grossman through set theory, in
flexible mine design, in Proceedings Orebody Modelling and
Strategic Mine Planning (ed: R Dimitrakopolous), pp 109-117 (The
R Dimitrakopolous), pp 323-328 (The Australasian Institute of
Hall, G, 2004. Multi-mine better than multiple mines, in Proceedings
Whittle, J, 2009. The global optimiser works – What next? in
Orebody Modelling and Strategic Mine Planning (ed: R
Proceedings Orebody Modelling and Strategic Mine Planning (ed: R
Dimitrakopolous), pp 3-5 (The Australasian Institute of Mining and
Lane, K F, 1988. The Economic Definition of Ore: Cut-off Grades in
Whittle, J and Wharton, C, 1995. Optimizing cut-off grades, in
Theory and Practice, 149 p (Mining Journal Books: London).
Proceedings 1995 Optimizing with Whittle: Strategic Mine Planning
Lerchs, H and Grossman, I F, 1965. Optimum design of open pit mines, (Whittle Programming Pty Ltd: Melbourne).
CIM Bulletin, volume 68 (Canadian Institute of Mining and
Metallurgy).
HOME
Integrated Strategic Planning at EKATI Diamond Mine

B Coutts1
ABSTRACT 5 Year Plan (5YP) and 2 Year Budget processes. This paper will
focus on the RDP which is the key strategic element of the
Effective strategic planning at EKATI Diamond Mine, Northwest
Territories, Canada, requires both the engagement of the business into the planning cycle. The RDP applies inclusive, divergent thinking at
process and the integration of the results into deliverable schedules and the directional phase of the planning cycle to understand the
developments. The goal is to position EKATI as the lowest cost premier Asset’s full potential value. Key sources of value during the RDP
producer of diamonds in the developed world. This will be accomplished process are realised through the integration of strategic direction
by maximising EKATI’s economic and environmental sustainability with physical input ranges, an investigation of value and business
through identifying and developing options with the potential to deliver a opportunities through unconstrained thinking and the provision
robust, long life, high quality asset. To achieve this, strategic planning must of a development plan to maximise full potential return at Asset
integrate directional policies with the physical input ranges applicable at level, including the value of exercising growth options. This
EKATI via an annual, comprehensive, inter-linked planning system to both process enables sound decisions through structured analysis of
optimise and challenge the present state. Diamonds present unique
the facts and facilitates the evaluation of uncertainty and its
uncertainties to the planning process. The evaluation of variance in
certainty and range within the multiple inputs enables the assessment of resultant impacts upon growth options.
strategic opportunities and risks. The development and application of an The effective communication and integration of strategic
EKATI specific evaluation and optimisation model facilitates the rapid, planning results within the planning cycle is achieved through
objective application and assessment of multiple strategic policies, structured output which includes clearly identified and
constraints and development options. Effective communication and constrained go-forward frameworks, approved by CSG
integration of strategic planning results into deliverable operational management, which form the basis for further evaluation during
schedules is critical to both maximise the value of EKATI and execute the planning cycle. Additional deliverables are a comprehensive
opportunities to grow a long-term sustainable business.
and prioritised list of development projects to move towards full
asset value and the identification of potential development and
INTRODUCTION improvement options including trigger values to move into their
execution.
BHP Billiton Diamonds Inc is the operator of the EKATI
Diamond Mine, which is located approximately 300 km Approval for the development of the EKATI mine was justified
northeast of Yellowknife and 200 km south of the Arctic Circle on the basis of the 1996 Feasibility Study (BHP Diamonds Inc,
in the Northwest Territories of Canada. EKATI’s environment is 1996). Since that seminal assessment, regular re-evaluations of
sub-Arctic desert, where the average annual daytime high is -9°C the EKATI strategic plan have provided updated economic
and permafrost occurs to depths of 330 m (Figure 1). The mine parameters upon which future business decisions have
operates as a fly-in, fly-out camp with 640 rooms and often >800 been made. Incorporated into these re-evaluations are changes
people on-site. The major route for annual re-supply and moving to multiple physical parameters, increases in operational
heavy loads is the seasonal winter road, operated over frozen experience, a dynamic financial environment, paradigm shifts in
lakes for approximately two months a year and shared with two marketing and the constant evolution of modelling tools. As such
other diamond mines. Following the initial discovery of the this paper presents a selective snapshot of a continuous process.
Lac de Gras diamondiferous kimberlites in 1991, over 150
kimberlites have been identified within the EKATI lease INPUTS
holdings. Since opening in 1998, EKATI has produced in excess
of 40 Mct of diamonds (to end FY08) from a combination of The key strategic imperative for EKATI is to unlock its long-term
open pit and underground mining of six kimberlite pipes. This potential in a manner that optimises the return to all stakeholders.
paper is a case study of the current strategic planning processes This can be accomplished by maximising EKATI’s economic
and systems at EKATI, the intended future developments to these and environmental sustainability through identifying and
systems and the effective wider business integration of both developing options with the potential to deliver a robust, long
inputs and results. life, high quality asset. To achieve this, the RDP must integrate
directional strategy with the physical input ranges applicable at
The planning cycle applied at EKATI is a BHP Billiton EKATI.
mandated procedure which allows for aligned planning
throughout the company. This procedure is mandated by Group,
delivered by the Customer Sector Group (CSG) and executed by Directional strategy
the Asset. In this case the CSG is Diamond and Specialty A clear and thorough determination of the present state of
Products (D&SP) and the Asset is EKATI. The planning cycle is EKATI is a critical component in the preparation of the RDP.
an annual integrated process that links high level directional Processes undertaken include performing a constraints analysis,
planning with detailed delivery planning (BHP Billiton, 2008). It conducting a value driver analysis, reviewing EKATI’s risk
starts with the Chief Executive Officer’s (CEO) Message, which register and updating safety, environmental and regulatory
reinforces the elements supporting the directional strategy. The requirements. Constraints analysis is a structured approach to the
directional planning phase considers full potential, key strategic analysis of constraints or ‘bottlenecks’ on the output of the
choices and imperatives evaluated through the Resource operation. Value driver analysis is conducted in order to
Development Potential (RDP) and Life of Asset (LOA) planning determine the key variables which affect the value of the
procedures. The delivery planning phase focuses on resource operation. The risk register lists identified and understood risks
allocation improvement initiatives and action and includes the to the entity. The active engagement of EKATI’s management
team is fundamental to achieving consensus in these analyses.
1. Superintendent Long Range Planning and Projects, BHP Billiton Development of EKATI’s required forward looking strategic
Diamonds Inc, 1102 - 4920 52nd Street, Yellowknife NT X1A-3T1, framework is achieved through the collation of all available
Canada. Email: [email protected] strategic inputs including the CEO message, the CSG strategic
B COUTTS
FIG 1 - EKATI location.
framework (Figure 2) and scenarios. Strategic scenarios are Physical input

plausible, widely divergent but internally consistent models of
Current strategic planning at EKATI optimises a mineral inventory
possible future worlds. Each may include a set of future macro
of 12 physically discrete and widely dispersed kimberlite
economic conditions and an assessment of likely implications as orebodies developed by open pit and/or underground operations.
to the market, costs, competitors and environmental issues. Each kimberlite pipe exhibits unique value and recovery
Review of these materials by EKATI’s management team characteristics in combination with specific capital and operating
generates the scope of the RDP. Essential to the directional requirements. Blended feed is processed in a single plant with
preparation is the collection and review of current Research and multiple potential operating profiles. All the physical inputs
Development (R&D) initiatives that could have a future impact evaluated in the RDP are subsequently applied in further
upon EKATI. Current R&D initiatives under evaluation at progressive stages in the planning cycle. Ensuring consistency of
EKATI include potential alternate methodologies for mining, value in these inputs is fundamental to achieving accurate and
energy and transport. Given the fact that EKATI’s product is sold comparable results from the planning cycle. Achieving this
into a complex and opaque global market, the direct involvement requires constant vigilance and communication within and
of marketing is vital to enable the assessment and review of between the many technical teams involved. Currently EKATI
market trends, forces driving change, marketing innovations and seeks to regulate this through controlled sign-off sheets of required
product concepts. As EKATI’s RDP must incorporate the full values. Development of an integrated planning database intended
potential mineral inventory, consultation with the exploration and to further facilitate both the accurate generation and secure
resource development group ensures the inclusion of all potential provision of input values is discussed further in the progression
greenfield and brownfield exploration results. Amalgamation of section.
the directional and physical elements is achieved through a
facilitated problem framing workshop. A problem framing Resources
workshop commences with the current state, clearly and As discussed in Dyck et al (2004) and Kleingeld and Nicholas
succinctly states the problem to be resolved, applies structured (2007) the modelling of diamond deposits is an iterative process of
decision hierarchies to generate both strategic decisions within analysis and interpretation, seeking to both identify and quantify
potential settings and divergent consistent themes for further critical aspects of risk and uncertainty to acceptable levels as
evaluation. The workshop participants ideally include the defined in Yates and Hodson (2006). The major elements to
broadest possible range of contributors of information, resource modelling at EKATI include diamond drilling for pipe
decision-makers and output owners. Participation of EKATI’s volume and internal geology which often occurs contemp-
management team is essential to ensure the resultant strategic oraneously with multiple phases of sampling for grade distribution
themes, decisions and alternatives provide an accurate and and diamond valuation. This enables both the deterministic and
comprehensive direction to the RDP. geostatistical modelling of kimblerlite volume and domains, and
INTEGRATED STRATEGIC PLANNING AT EKATI DIAMOND MINE
Strategic
Pillars Strategic initiatives
imperative
Right 1. Relentless pursuit of Zero Harm to People

People,
Right Place 2. Employer of Choice
Maximize
Maximise 3. Relentless pursuit of Zero Harm to the environment
Licence to
operate
EKATI’s full 4. Operator of choice for communities and government
potential
5. Identify and optimize growth options
World Class
Asset
6. Creation of a cost conscious culture
7. Improve operating discipline
FIG 2 - EKATI strategy tree.
the simulation of diamond grade and calculation of resource Each resource has a unique ore revenue (US$/tonne) which is
ranges at various mining scales. The classification of the resource generated from the average recovered grade (carats per tonne)
uses a combination of probabilistic and deterministic methods. It multiplied by average diamond value (US$/carat). Average
is important to understand the internal geology of the kimberlite diamond value (diamond reference value) is estimated for each
pipe before undertaking any modelling exercise as internal phases pipe or some cases multiple geological domains within a pipe
may act as either hard or soft boundaries for modelling. Trends using exploration or production parcels ranging in size from
within and across domains may also play a key role due to both several hundred carats to tens of thousands of carats. The diamond
the volcanic and sedimentary process involved in kimberlite value estimate is relatively complicated compared to most
emplacement. Harrison et al, 2009 discuss the creation of commodities. Diamonds occur in a skewed distribution of sizes
uncertainty based spatial models at EKATI as shown in Figure 3 and qualities with a price variation up to three orders of magnitude
using sequential Gaussian simulation (SGS) methods detailed in for a single diamond size. Diamonds within a kimberlite range in
Goovaerts (1997). quality from very low value boart (fibrous diamond) to very high
gem quality stones. More than 2200 categories comprise the
current EKATI diamond price book (PB). Average diamond value
is a function of diamond size distribution and diamond quality.
The highest value populations have both a coarse size distribution
and high proportion of high quality stones. The P50 diamond
reference value is a function of diamond size distribution, diamond
quality classification per size category (price point) and process
plant recovery. Uncertainty associated with diamond value
estimation is related directly to parcel size assuming a detailed,
well constrained value sort. This relationship is presented in
Figure 4, along with an indicative cost associated with bulk
sampling at EKATI. The uncertainty due to parcel size can be
estimated using Monte Carlo random testing of very large
production parcels. The ideal parcel size for commercial
kimberlite evaluation is approximately 5000 carats. However,
there are cost and time constraints in obtaining very large diamond
parcels at EKATI due to the difficult setting of the pipes (typically
FIG 3 - Sequential Gaussian Simulation (SGS) estimation of

3
carats/m , Jay Pipe, EKATI Diamond Mine (Harrison et al, 2009). FIG 4 - Parcel size, uncertainty and indicative sampling cost.
B COUTTS
situated in lakes with less than ten metres of glacial overburden). verified though auditing of plant reject material. Incorporation
The evaluation is sequential due to multiple pipe development and of the current and correct plant operating strategy is a crucial input
seasonal constraints (winter bulk sampling programs). At EKATI to the RDP. Assessment of variations to this key parameter
those pipes in production have very large parcels resulting from is a critical element to ensure optimisation of the screen-recovery
large scale production tests and the corresponding diamond value -value relationship for a specific ore blend.
estimates are very well constrained. Pipes still in the process of
investment evaluation studies have smaller, variably sized, parcels Financial
with proportionate value uncertainty. The revenue impact of parcel
size risk can be positive or negative. The aim of increasing the size Prior to any work being carried out in the RDP it is important to
of the diamond parcel through collection and processing of bulk define the costs used. A cost model combining both short-term
samples is to reduce the variability of the range of possible and long-term budgets is used as the baseline for planned future
diamond values. The valuation of the diamond parcels is performance with evaluation and redistribution of costs undertaken
periodically re-fixed based on the prices achieved from the market to ensure the correct costs are applied. Operating and capital cost
and this ‘price’ is called the price book. Prices are regularly estimates applied in the RDP are sourced from either budget
estimates or relevant engineering studies. Unit costs are
reviewed but for planning and market forecast purposes the Price
determined for open pit and underground mining and the process
Book is fixed and the commodity price forecast (Diamond Price plant. Recognising the granularity of the analysis, a distinction is
Index) is applied to the fixed book price. EKATI’s product is not made within unit costs between variable and fixed costs. Fixed
sold continuously but in discrete events termed Diamond Sales. general and administrative costs comprise a large component of
Each time the price book is updated it is numbered according to the EKATI cost model, reflecting the expense of maintaining the
the diamond sales cycle to which it relates. footprint required to operate in a remote and challenging
EKATI’s RDP is all inclusive, evaluating the operations full environment. All costs are in Canadian dollars on 100 per cent
risked mineral inventory which includes measured, indicated and terms and are exchanged to US dollars (US$) to be comparable to
inferred resources as well as a full range of exploration, regional revenue which is in US$. The active engagement of EKATI’s
and near-mine potential resources. The application of resource management team is fundamental to achieving consensus in the
range analysis generates an estimate of potential ranges within accuracy of costs generated and applied. The value impact of cost
the risked mineral inventory, and assigns probability weightings variations is an important evaluation within the RDP. Range
to the ranges. At EKATI these ranges often incorporate potential analysis will be greatly facilitated by the development of an
geologic phase based grade changes, and volumetrically minor integrated cost modelling system, discussed further in the
satellite kimberlites. The issue of technical certainty is managed progression section.
through planning cases built by the incremental addition of
resources based on the status of specific resource knowledge Environmental
and/or design maturity.
Reclamation and closure costs are sourced from a detailed closure
Engineering cost model. Working from the detailed cost estimating in the
model, pipe specific closure costs are allocated accordingly,
Resource schedules are applied in the RDP in a range of
including those pipes where development has not yet commenced.
formats and obtained from a range of sources, including
underground production scheduled at a fixed rate in all cases as Where an underground operation is currently in production or
provided by the underground operations or relevant engineering planned under an existing open pit, the open pit closure costs are
studies, reserves on a bench by bench basis from designed pits, allocated to the underground operation. Pipe specific reclamation
scheduled tonnes on a bench by bench basis from Whittle shells costs are applied at the same time as pipe development costs. The
and Stockpiles. Schedules are updated as required to reflect remainder (and far greater proportion) of the closure costs is
either current Whittle optimisations or the re-design of scheduled at the end of the mine life for each scenario. The timing
operational mines. of the requirement to commit closure funds impacts the optimised
Additional operating parameter assumptions that are generated mining sequence determined and additional information in this
for the RDP include haul cycle times, haulage costs, incremental area will benefit future plans. Pipe specific ‘commencement of
bench costs, sinking rates, load factors and equivalent standard production’ limitations are applied to reflect constraints associated
truck factors. with project lead times resulting from environmental baseline
studies, permitting, regulatory approvals and engineering studies.
Processing
Diamonds occur as discrete particles within a kimberlite pipe and
Marketing
must be recovered as intact entities to be of commercial value. Rough diamonds are not a commodity in the traditional sense
Given the low concentration of diamond in kimberlite ore – one to due to a lack of benchmark trading prices, a lack of homogeneity
two parts per 20 000 000 on average worldwide, one to two parts in product characteristics/value, a lack of transparency in trading
per 10 000 000 at EKATI – large volumes of material must be due to proprietary producer-controlled marketing systems, and a
processed to obtain economical quantities of product (ie lack of transparency and homogeneity in value-add from rough
diamonds). Mineral processing at EKATI is based on the unique to polished to retail. Additionally, there is no established forward
physical properties of diamond that allow for standard reduction market. The gem diamond market has historically been supply
and recovery by means of three main unit operations which are the constrained due to previous monopolistic controls, and demand
crushing and scrubbing of kimberlite ore followed by primary driven by emotional marketing. Since the commencement of
concentration by heavy medium separation and then secondary EKATI several key developments have resulted in a transition to
concentration by X-ray sorting. All of these operations are an open market characterised by a long-term scarcity of supply.
performed on site. Final diamond cleaning and sorting is These developments include the emergence of multiple
completed at an off-site sorting and valuation facility. Screens of competitive sources of supply, the abandonment of the market
specific size which are applied during the first two operations custodian role by the incumbent and the liquidation of diamond
control the recovery rate of diamonds from the material flow. stockpiles. In addition, exploration has not succeeded in finding
There are specific relationships between the screen size, plant substantial new resources while mature mines face declining
throughput rates and the value of the recovered product. production. Demand is correlated with gross domestic product
Reconciliation and validation of this recovery relationship is (GDP). BHP Billiton’s diamond price index (DPI) is a
long-range forecast in real terms of relative market value to a permissible option (eg an open pit) is not scheduled by COMET
fixed price book. It is applied to diamond valuations (US$/carat) this is an indication that this option is not value additive under
anchored to the same price book to forecast revenue in real US the applied conditions. For a fuller discussion on the algorithms
dollars. The Diamond Price Protocol includes P99, P90, P50, P10 and assumptions applied within COMET the reader is referred to
and P01 DPI forecasts that reflect the uncertainty range in Wooller (2007) and additional discussion on the integrated
long-term price trends. applications of the software is provided by King (2007).
The COMET model currently applied at EKATI is based on
PROCESS three Excel workbooks to allow the generation and control of both
input and output in a useable form. Figure 5 is a high level
overview of the COMET model. Specific elements within the
Tools
Inputs workbook include unique names for each resource and
The EKATI Strategic planning process utilises a variety of design, ensuring no duplication. Sequencing and naming of
industry standard tools. These tools are summarised in this section, resources and designs is critical, as each worksheet contains
with greater detail provided regarding the application of the substantial resource and design specific data referenced
COMET optimisation software. Currently both short-term and downstream in COMET. Data within each worksheet includes ore
long-term scheduling is undertaken using a combination of tonnes and grade by phase, waste tonnes by rock type, dilution,
Microsoft Excel worksheets and Runge XPAC databases, recovery and moisture factors, truck productivity and mining rates.
dependent upon the origin and application of the data. A project is As the COMET optimisation process applies an identical
underway with the intent to transition all scheduling into XPAC. weighting and confidence to the evaluation of all individual
Pit optimisation is performed using the Gemcom Whittle resources, resource uncertainty is managed through the
optimisation tool. Pit shells are re-assessed as required dependent incremental assessment and evaluation of specific resource
upon changes in the key input values. Diamond values applied for packages of similar confidence. The Phase Generation File
Whittle optimisation are those forecast to apply at the completion workbook converts resource, physical and financial inputs into
of mining. Assessment of Whittle results for various economic format required by COMET. The scenario control worksheet
scenarios indicate that the P50 design captures almost all of both enables macro driven generation of required files. The workbook
P10 and P90 kimberlite tonnage. This apparent lack of sensitivity populates phase (resources) and schedule (physicals and
to diamond value in pit optimisation reflects the discrete ore-waste financials) folders. Data within the workbook includes costs
character of the kimberlite pipes. Given the relatively small size of (operating, capital and closure), processing scenarios and
most of EKATI’s pipes, phase optimisation is only undertaken on economic assumptions. The COMET GUI contains specified
specific, larger resources. Pit design is completed utilising constraints, policies and schedules to enable the requested
Maptek’s Vulcan software suite. optimisation to occur. The GUI contains the Optimisation Engine,
Prior to 2003, the optimisation of EKATI’s value was a macro for multiple scenarios and other control features. It
undertaken through Excel based Discounted Cash Flow models. generates the physical and financial outputs for validation and
Strategic optimisation and options at EKATI are currently reporting. The resultant NPV is reported on a 100 per cent real
evaluated through STOPS COMET software, both as a tool pretax $US in country basis. To enable further validation,
within the RDP and for undertaking specific evaluations as evaluation and reporting, the results of all COMET runs are
requested by management. The development and application of processed through a macro driven workbook within Pivot tables.
an EKATI-specific COMET model enabled the objective The nature of the optimisation process and applied constraints can
application and assessment of multiple policies, constraints and result in operational schedules which appear impractical as shown
options in a timely manner. COMET is an optimisation program in Figure 6. This is acceptable as the purpose of COMET is to
that is used to run alternative life-of-mine scenarios in a optimise the order of extraction. Further work on the maximised
relatively short period of time. It is a broad brush analysis tool NPV schedule during the next stage in the planning cycle
which is useful for making informed decisions on which options generates a deliverable mining schedule. Additional strategic and
warrant more detailed consideration. COMET is a software operational applications of COMET at EKATI include both the
package that iterates a mining sequence within a set of assessment of constraint specific ‘trigger values’ for options and
predetermined constraints to give a resultant sequence which projects, and the evaluation of the impacts of dilution (in situ and
generates the maximum Net Present Value (NPV). If a mining) and cost benefits of control programs.
Input Workbook
Data: Physical Resources
Comet GUI Workbook

Data: Constraints, Policies, Schedules
Functions: Control Feature,
Optimization Engine and Output
Phase File Generation

Workbook Generated .txt Phase Files
Conversion
Data: Costs, Assumptions Data: Physical Resources,
Functions: Case Setup and Costs, Assumptions
conversion of inputs into txt files
FIG 5 - Schematic of EKATI COMET model.
B COUTTS
TOTAL MATERIAL MOVEMENT

50
45
40
35
MILLIONS TONNES
30
25
20
15
10
0
20 11
2 015
2 019
202 0
202 4
20 28
2008
2009
2010
2012
2013
2014
2016
2017
2018
2021
2022
2023
2025
2026
2027
2029
2030
FIG 6 - Representative COMET schedule.
Risks and opportunities That COMET routinely selects this option when it is made
available highlights the potential of a plant expansion to add value.
The core building block within the EKATI strategic planning Greater certainty is required on costs and throughput rates for the
process is the present state, or the base case. Constraints applied to plant expansion option. A hypothetical exploration pipe has been
the base case are the current and scheduled physical, operational, previously included to evaluate the scheduling and value impacts
economic and directional parameters. It also includes projects in of an exploration discovery. This emphasised the significant value
active execution. Considerable verification and focus is applied to to EKATI of ongoing exploration aimed at such a discovery.
ensure the base case is fully and accurately represented and At EKATI both the base case and selected development options
modelled within the process tools. At EKATI the resultant NPV are stress tested for their sensitivity to a range of potential
from COMET evaluation of the base case is set at 100 per cent. The economic and operational scenarios. This includes an assessment
results from all additional strategic assessment are always reported on any resultant changes to resource scheduling or developmental
as a percentage relative to the base case. The establishment of the priorities. These scenarios (either singularly or as combinations)
base case and the functionality of the COMET tool permit the include high/low risk production applying a variance in process
incremental assessment of development projects, tactical and plant throughput and reflecting poor winter road seasons or other
strategic options and alternate scenarios. Contingency and operational restrictions. Movements in DPI and value are assessed,
sensitivity assessments facilitate the evaluation and understanding of potentially caused either by new mines not coming on line or
ranges (P99 through to P01) and certainties, upon both the base case substantially weakened global demand. Potential variance in the
and further incremental scenarios. Variation of single constraints scheduled production rates of high value ore sources, reflecting
within the model enables the identification and quantification of key either strong operational performance or operational restrictions, is
value drivers, opportunities for improvement and potential risks to evaluated. Ranges in cost profiles, reflecting either success in cost
project value. The incremental value assessment of development reduction programs or increase in direct or indirect costs (fuel, etc)
projects enables the relative prioritisation of required investment are included in the scenarios. The potential application of mining
studies. The quantification of the inherent certainty within each methods alternative to conventional drill and blast for production
development project provides an understanding of the associated scale mining at specific kimberlite pipes is assessed. In addition,
risk. As a progression of this, the variation of single constraints for the impacts of changes in the processing plant operational
specific development projects enables the identification and parameters of throughput, screen size and value recovery,
quantification of key project drivers such as operating costs, capital including any required capital spend is evaluated.
costs or diamond value. The progressive building of project
groupings also permits an understanding of the impact to EKATI’s
value from changes to the project development schedules, again DELIVERY
facilitating the focus of internal energy upon the required projects. Effective communication and integration of value driven
Critical elements of option assessment include identifying the strategic planning results within the BHP Billiton planning cycle
key drivers for the preservation of options, and the trigger values is critical to maximise the value of EKATI and develop
that will result in the activation of those options. At EKATI, opportunities for a safe, long-term sustainable business.
options considered and preserved include several large tonnage,
low margin resources which despite being un-economic under
Integration
current conditions are EKATI’s key drivers of scale. The trigger
values for these options are now the major goal of dedicated Evaluation of EKATI’s base case, risks and opportunities
Business Excellence and R&D programs. Plant expansions and the provides a series of quantified metrics that assist management
resultant opportunity to reduce overall unit costs are also assessed. to determine the appropriate ‘go-forward’ strategic planning
options. The key metric used is the relative NPV. Other metrics strategic plan are selectively shared with the wider stakeholder
considered include the capital efficiency ratio (CER), the life of community, to both emphasis development opportunities for
the mining operation, total scheduled tonnes and the technical EKATI as a safe, long-term sustainable business and to seek to
uncertainty of development projects and options. A review of all proactively engage community expectations.
evaluations is held with the EKATI management team with the
purpose of verifying the results of the RDP process and PROGRESSION
providing formal sign off on the go forward case(s) for the life of
asset (LOA) scheduling. Included in the recommendation for the Over the preceding 13 years strategic planning at EKATI has
LOA schedules is which schedule will become the basis for the constantly evolved. The current process delivers value focused
5YP. The recommended ‘go-forward’ planning cases must integration throughout the planning cycle, from directional to
comply with specific JORC related and technical certainty delivery. Future improvements to the process, some currently in
criteria. Two specific planning cases are recommended for execution, are intended to proactively progress this ongoing
further assessment in the LOA. These are the LOA ‘Optimised evolution.
Without’ – the Base Case with projects and options in execution,
and the LOA ‘Optimised With’ – the above case with additional
value positive development projects and options. The go forward
Inputs
cases from the RDP are approved by the CSG and therefore form A potential development is the imposition of a carbon tax in
the basis of the LOA, 5YP and ultimately the 2 Year Budgets. Canada. Given EKATI’s current full dependence upon diesel for
The go forward cases also lay the groundwork for the Portfolio supply, heat, fleet and power, this tax would have a significant
Valuation Review (PVR), which is the formal internal valuation impact upon both the base case and development projects. The
of the Asset. COMET provides an overall sequence through the sensitivity assessment of this on selected planning cases would
development options to maximise value. This sequence is used as quantify both the costs and benefits of potential mitigation
the basis for the PVR. These cases are scheduled as part of the activities such as the development of alternate energy sources. A
LOA process in XPAC for consistency of approach. significant drawback in the current process methodology is the
The planning cycle is an annual process, reflecting the reality of de-facto imposition of annual production targets as constraints,
operating and developing under constantly changing and evolving through the ‘fixing’ of resource development to honour
physical, financial, regulatory and strategic parameters. At EKATI scheduled tonnes. This removes the ability of the optimisation
major changes in both short and long-term planning has been software to generate a truly NPV maximised result, and
driven by variations in the certainty of resource elements such as unconstrained analysis is critical to evaluate and understand key
tonnage, grade and diamond valuation as a result of continued NPV drivers. Given that EKATI is a mill constrained operation,
sampling. Additional changes reflect the increasing understanding the assessment of potential stockpiling strategies is warranted.
of the key drivers of plant throughput and recovery. Ongoing This has not been fully evaluated to date, in part due to the
engineering studies result in changes to both the scale and historical acceptance of production targets as constraints.
certainty of operating and capital cost estimates for development EKATI’s unique geographical and physical parameters have
projects. Additional planning cases are reviewed and approved by required innovative solutions. Current R&D programs into
management for assessment in the LOA, both ensuring the alternate methods of mining, ore haulage, transportation and
preservation of options and to enable the most complete long energy continue this tradition. The integration of R&D into
range view of the development potential of EKATI. strategic planning enables both evaluation and prioritisation of
opportunities. Future large scale development projects at EKATI
Communication are low margin. One of the solutions to this challenge is the
selective sequencing and mining of higher grade zones within
The integration of the RDP ensures formal communication of the these resources. Evaluation of the resultant impacts may provide
principal strategic planning case into EKATI’s business plan. justification for the development of required supporting
During all reviews significant use is made of graphical
programs such as grade control.
presentation of data and results, including tornado, waterfall and
bubble charts. A review of all evaluations is held with both
the EKATI management team and the EKATI operational Process
leadership team to discuss the results of the RDP process. An integrated planning system currently being developed is
Recommendations, with supporting justification, are provided
based on an XPAC-XERAS scheduling and budgeting core with
based on assessments undertaken during the process. These
programmed links to COMET and Whittle optimisation. It will
recommendations include the confirmation of EKATI’s key
facilitate both the consistent secure provision of input values and
business drivers and the required trigger values to enable future
options. They address operational improvements, such as the iterative assessment of development options. The system will
changes to the processing configuration or confirmation of the enable alignment between the many suppliers and customers of
economic viability of alternate mining methods. Timelines for the contained data. The development of pipe-specific throughput
permitting and/or environmental base line studies are factors will enable the application of time factored cash flow
recommended to ensure deep preparation for supporting project grades ($/hour) at EKATI, further enhancing both directional and
development schedules. The requirements for additional drilling deliverable planning and scheduling. Given that specific
of development projects to further manage certainty of key elements within EKATI’s planning process contain inherent
project specific variables such as grade, volume, hydrology or highly variable uncertainties, the application of real options
structure is confirmed. Modifications are recommended to Analysis (ROA) to both increase the explicit understanding of
EKATI’s mobile or fixed plant, such as future mobile fleet underlying assumptions, and to better predict and understand the
requirements or changes in power generation to support required value of project flexibility must be assessed. A real option is the
infrastructure. Predictions of manning are generated based on the right to undertake a business decision and ROA applies put
future operational profile. The potential impacts for fine scale pit option and call option valuation techniques to these decisions.
design to increase reserves is highlighted and recommend future
specific design review work, including the assessment of split Delivery
shell designs in larger pits. The sensitivity on plant throughput
and pipe specific margins is provided to support the case for It is an ongoing obligation of strategic planning at EKATI (and
development of both throughput data in resource models and elsewhere) to challenge accepted paradigms. Engagement in such
grade control programs. Specific elements of the EKATI discussion is most effectively achieved by ‘Size of the Prize’
B COUTTS
assessment of the potential options. For example, what are the Dyck, D R, Oshust, P O, Carlson, J A, Nowicki, T E and Mullins, M P,
resultant increases in both project value and life from achieving a 2004. Effective resource estimates for primary diamond deposits
from the EKATI Diamond Mine, Canada, in Proceedings Eighth
specific and sustainable unit operating cost? A key element in
International Kimberlite Conference Selected Papers (eds: R H
EKATI’s long-term sustainability is proactive engagement of Mitchell, H S Gruetter, L M Heaman, B H Scott Smith and
regulators and the wider community in the development of fully T Stachel), vol 1, pp 317-335.
detailed and cost inclusive Closure Plans. The integration and Goovaerts, P, 1997. Geostatistics for Natural Resources Evaluation,
evaluation of closure within strategic planning is critical. 483 p (Oxford University Press: New York).
Harrison, S, Leuangthong, O, Crawford, B and Oshust, P, 2009.
CONCLUSIONS Uncertainty based grade modelling of kimberlite: A case study of the
Jay kimberlite pipe, EKATI Diamond Mine, Canada, in Proceedings
Effective strategic planning at EKATI requires both the Ninth International Kimberlite Conference.
engagement of the business into the process, and the integration of King, B, 2007. Integrated strategy optimisation for complex operations,
the results into deliverable schedules and developments. This is in Orebody Modelling and Strategic Mine Planning, second edition
achieved through an inclusive review of strategic direction and an (ed: R Dimitrakopoulos), pp 31-35 (The Australasian Institute of
annual, comprehensive, inter-linked planning system to both Mining and Metallurgy: Melbourne).
optimise and challenge the present state. Integrated strategic Kleingeld, W J and Nicholas, G D, 2007. Diamond resources and
reserves – Technical uncertainties affecting their estimation,
planning will be a key element in successfully realising EKATI’s classification and valuation, in Orebody Modelling and Strategic
potential as the lowest cost premier producer of diamonds in the Mine Planning, second edition (ed: R Dimitrakopoulos), pp 207-213
developed world. (The Australasian Institute of Mining and Metallurgy: Melbourne).
Wooller, R, 2007. Optimising multiple operating policies for exploiting
ACKNOWLEDGEMENT complex resources – An overview of the COMET scheduler, in
BHP Billiton Diamonds Inc management are thanked for the R Dimitrakopoulos), pp 289-296 (The Australasian Institute of
permission to publish this paper. This paper presents a selective Mining and Metallurgy: Melbourne).
snapshot of the continuous process of strategic planning at Yeates, G and Hodson, D, 2006. Resource classification – Keeping the
EKATI and the contribution of all BHP Billiton staff and end in sight, in Proceedings Sixth International Mining Geology
Conference, pp 97-104 (The Australasian Institute of Mining and
consultants is most warmly recognised. Metallurgy: Melbourne).
REFERENCES
BHP Billiton, 2008. Practioner’s guide to planning.
BHP Diamonds Inc, 1996. NWT diamonds project feasibility report.
HOME
Stochastic Long-Term Production Scheduling of Iron Ore

Deposits – Integrating Joint Multi-Element Geological
Uncertainty
J Benndorf1 and R Dimitrakopoulos2
ABSTRACT 1999; Dimitrakopoulos, 2007). Based on drill hole data and their
Meeting production targets in terms of ore quantity and quality is critical
statistical properties, conditional simulations generate several
for a successful mining operation. In situ grade variability and uncertainty equally probable models (or scenarios) of a deposit, each
about the spatial distribution of ore and quality parameter cause both reproducing available data and information, statistics and spatial
deviations from production targets and general financial deficits. A continuity, that is, the in situ variability of the data. The difference
stochastic integer programming formulation (SIP) is developed herein to between the equally probably scenarios are a quantitative
integrate geological uncertainty described by sets of equally possible measure/description of uncertainty. The subsequent integration of
scenarios of the unknown orebody. The SIP formulation accounts not only this grade uncertainty and local variability into mine planning
for discounted cash flows and deviations from production targets, but also optimisation allows for the understanding and control of
discounts geological risk, while accounting for practical mining.
Application at an iron ore deposit in Western Australia shows the ability of geological risk. This in turn aims to decrease project risk and
the approach to control risk of deviating from production targets over time. increase profitability.
Comparison shows that the stochastically generated mine plan exhibits less The detrimental effects to mine planning optimisation from
risk in deviating from quality targets than the traditional mine planning ignoring in situ grade variability and uncertainty in the
approach based on a single interpolated orebody model. description of orebodies are well documented (Ravenscroft,
1992; Dowd 1997; Dimitrakopoulos, Farrelly and Godoy, 2002;
and others). For example, Dimitrakopoulos, Farrelly and Godoy
INTRODUCTION (2002) show the danger of relying on estimated (average type)
Long-term mine planning and production scheduling aim to define orebody models when optimising. In their example, net present
the ‘best’ mine plan subject to the constraints imposed by physical value (NPV) assessment of the conventionally generated
and geological conditions, policies and the operational mining life-of-mine schedule using simulated scenarios of the orebody
approach. The term ‘best’ is defined by management objectives. shows the most likely NPV to be materialised standing at 25 per
These typically include maximising the monetary value of the cent lower than forecasted. The substantially positive
mining project as well as meeting customer expectations and contribution of accounting for grade uncertainty through multiple
guaranteeing a safe operation. The expectations of customers are simulated scenarios and new stochastic optimisation approaches
defined largely in terms of ore tonnage and ore quality is also well documented. Godoy and Dimitrakopoulos (2004)
characteristics to be delivered. In the case of multi-element show a long-term production scheduling approach based on
deposits, ore quality characteristics are defined by multiple inter- simulated annealing applied to a gold mine to result in a 28 per
correlated elements. For example, in iron ore deposits, the cent increase of project value compared to the conventional
elements iron (Fe), phosphorus (P), silica (SiO2), alumina (Al2O3) approach. Leite and Dimitrakopoulos (2007) show the same
order of improvement using this approach at a copper deposit. A
and loss of ignition (LOI) are critical for ore quality. Additionally,
more general and flexible long-term production scheduling
in many cases ore is produced out of multiple pits with different
method that allows the control of geological risk between
ore characteristics. The goal of any global, long-term mine production periods in terms of magnitude and variability is based
planning approach is to send the most homogeneous ore blend out on stochastic integer programming or SIP (Birge and Louveaux,
of multiple pits, meeting customer specifications, while 1997) and it is documented in Ramazan and Dimitrakopoulos
guaranteeing optimal pit development and maximising the utilis- (2007b) and Leite and Dimitrakopoulos (2010, in this volume).
ation of available mineral resources. In practice, however, when An application of the SIP formulation to the long-term
implementing a mine plan, differences frequently occur between production scheduling of a single-element deposit demonstrates
the produced ore quantity and quality characteristics. It is well its effectiveness and advantages in terms of additional project
recognised that uncertainty in the description of the spatial value and associated risk management even for a relatively short
distribution of grades of various pertinent elements in the orebody life-of-mine.
as well as their in situ variability are major contributors to these This paper contributes a mine planning optimisation approach
differences.
that addresses joint multi-element grade uncertainty, as common
Traditional approaches to mine planning optimisation are based in many mineral deposits, such as iron ore. More specifically, the
on a single estimated model of the orebody that is unable to stochastic integer programming approach of Ramazan and
account for in situ variability and uncertainty associated with the Dimitrakopoulos (2007b) is expanded to:
description of the orebody (David, 1977; David, 1988). Contrary
to estimation techniques, a different set of techniques provide a • multi-element deposits, and
tool to address shortcomings of estimation methods, termed • includes new mineability constraints to facilitate accessibility
conditional simulation (Goovaerts, 1997; Chiles and Delfiner, and equipment size constraints.
In addition, the formulation developed herein is exhaustively
tested in an application at an open pit iron ore mine in Western
1. GAusIMM, MIBRAG MBH, Germany. Australia and within the context of multi-pit production
Email: [email protected] planning. Testing includes the ability of the SIP to control the
2. FAusIMM, Professor and Director, COSMO – Stochastic Mine risk of deviating from production targets in terms of ore quality
Planning Laboratory, Department of Mining and Materials characteristics. In the next sections, the stochastic mathematical
Engineering, McGill University, Montreal QC H3A 2A7, Canada. programming formulation is first presented. The application and
Email: [email protected] testing of the formulation are presented.
J BENNDORF and R DIMITRAKOPOULOS
P K
STOCHASTIC PRODUCTION SCHEDULING − ∑ ∑ (c SM .Y1tj )
t=1 j=1
Global optimisation of long-term production scheduling
addresses issues of optimal sequencing considering multiple pits, where:
multiple elements, blending issues, stockpiling options and
alternative processing or product options (Whittle, 2007). The P is the number of periods
task of long-term production scheduling in a multi-pit operation N denotes the total number of blocks to schedule
can be divided into two stages. The first stage is a multi-pit
scheduling approach, which defines ultimate pit outlines as well S represents the number of simulated orebody models used to
as proportions and element qualities, where each pit and period capture geological uncertainty
contribute to the global target in order to optimise the global R is the number of targets including grade targets for different
asset. In the second stage the physical extraction sequence of elements and ore tonnage targets
blocks in each single pit is defined as constraints to production c ti represents the economic contribution of block number i
rates and targeted element grades implied by the multi-pit when mined in period t and is a representation of the
scheduling approach. This contribution concentrates on the expected economic value over all values of block i at time t
long-term scheduling of a single pit – multi-pit scheduling derived from each realisation s E{(NPV)ti }
approaches have already been successfully implemented, eg
BLASOR, developed in BHP Billiton’s Technology group (Stone x ti is a variable representing the percentage of block i mined in
et al, 2007). period t – if an x ti variable is defined as binary (0 or 1), it is
assigned 1 if block i is mined in period t and assigned 0 if
The goal of long-term production scheduling under grade
not
uncertainty of single pits is to define a physical extraction
sequence of blocks over periods so as to meet multiple goals. s
qu tr is the upper deviation from production target r at time t
These goals include: considering orebody model s
• best mine development and best use of available mineral yu tr is the unit cost of squrt to penalise excess production
resources for a maximisation of the monetary value of the s
ql tr is the lower deviation from production target r at time t
asset, considering orebody model s
• control of risk of deviating from production targets, and yl tr is the unit cost of sql tr to penalise a deficit in production
• guarantees of a safe operation.
Y1tj is the number of surrounding blocks, which are not mined
In this context, controlling the risk of deviating from production in the period t or earlier when mining block j
targets is a major contribution and involves controlling
Surrounding blocks are those, which are no more than three
probabilities and magnitudes of deviations from production
blocks apart in each direction (Figure 1). The costs cSM are
targets, as well as fluctuation of produced grades over periods. The
penalties associated with Y1tj . Note that this penalty only applies
underlying geological uncertainty is captured by a set of
to a subset K of all blocks N. To avoid overlapping, only every
conditionally simulated orebody models. Generally, production
third block in each direction is considered to be the central block j.
targets may be in terms of produced ore and waste tonnes and
grades of different elements. Constraints are in terms of The first part of the objective function is used for maximising
the discounted economic value in the context of the global
practicality of the schedule guaranteeing equipment accessibility,
optimisation. Note the global multi-pit approach accounts for
mining capacity, processing capacity, geotechnical aspects as well
interactions between different pits and aims to maximise usage of
as blending requirements.
Stochastic formulation for long-term production

scheduling
A general formulation for long-term production scheduling

under geological uncertainty for multi-element deposits based
on SIP is presented next. It is based on the single element
formulation in Ramazan and Dimitrakopoulos (2007b). The
objective function and relevant constraints are explained in
detail.
j
Objective function
The SIP objective function, presented here for scheduling

multi-element single deposits, combines several goals. It aims to
generate a production schedule that optimises the economic pit
development considering constraints imposed by the global
multi-pit approach, while minimising deviations from production
targets in terms of tonnages and ore-quality as well as minimising
costs of non-smooth mining. Equation 1 presents the three parts of
the objective function:
P N S P R
Maximise ∑ ∑ c ti .x ti -∑ ∑ ∑ ( s qu tr .yu tr + ql tr .yl tr ) (1) FIG 1 - Inner and outer window around block j in smooth mining
t = 1 i= 1 s = 1 t = 1 r= 1 constraint (after Dimitrakopoulos and Ramazan, 2004).
STOCHASTIC LONG-TERM PRODUCTION SCHEDULING OF IRON ORE DEPOSITS
resources and global value. The first part in Equation 1 maximises description of input data are discussed subsequently, in particular
the local NPV of the single pit under consideration aiming to the process of incorporating the stochastic production scheduling
define an optimal mine development constricted by the global approach of a single deposit into the global multi-pit scheduling
plan. It accounts for profit-defining aspects, such as stripping ratio. problem. The input in terms of simulated orebody models is
The discounted economic block value is calculated as expected presented as well as the operational, economical and risk
value from each realisation. The second part of the objective controlling parameters. Following, the practical approach of
function handles the deviations from production targets imposed scheduling Yandi Central 1 is detailed, including the practical
by the multi-pit scheduling approach for each simulated orebody
implementation of the scheduling formulation and the manual
model s including grades of all elements and ore tonnage. By
mine design to convert results to a practical schedule. A
optimising over S possible scenarios, captured through multiple
equally probable orebody models, this part of the objective comparison between schedules generated using a stochastic
function aims to control uncertainty and variability of the formulation to those using a deterministic formulation
produced grades and ore tonnage. The magnitude of grade considering one estimated orebody model is found at the end of
variability in the generated schedule is controlled for each element this section and demonstrates the benefit of the stochastic
e considered and time period t by penalties associated with approach.
deviations s ql tr and s qu tr . Note that deviations for each target and
period yu tr and yl tr are calculated by the corresponding constraints, Yandi operation and current production
which are the grade constraint and the ore tonnage constraint. Part scheduling practice
three of the objective function controls smooth mining by
penalising not mining adjacent blocks in same period, the central The Yandi Central 1 deposit is part of the larger Yandi channel
block j is scheduled, or earlier (Figure 1). Yl tj represents hereby iron deposits (CID), which occurs alongside the Marillana-
the percentage of the eight directly adjacent blocks and the 25 Yandicoognica Creek system about 120 km northwest of
blocks that are two block-widths distant, which have not been Newman, Western Australia. This deposit is part of the Yandi
mined in the same period as block j. Deviations of smooth mining joint venture operation, which includes multiple pits. The
for each considered block j and period t Yl tj are calculated in the fundamental objective of this complex operation is the achieve-
smooth mining constraint. The priorities of the three competing ment of customer defined on-grade shipments at lowest costs by
parts in the objective function are controlled by the magnitude of optimally blending from different pits with a diverse range of
corresponding cost parameters for each part relative to each other. resource grades. Critical geochemical parameters when evaluating
The mine planner has to adjust these parameters so to define the the deposit are iron content (Fe), silica content (SiO2), alumina
best schedule that compromises his objectives, for example the content (Al2O3), phosphorus content (P) and the water and organic
level of risk the planner is willing to accept. The related content measured as loss on ignition (LOI), as they influence the
constraints are detailed by Benndorf (2005). physical and chemical properties of the product and the
performance of the beneficiation process.
Controlling risk over time for different objectives For the global multi-pit optimisation of the Yandi joint venture
operation, BHP Billiton’s Technology group developed a
As presented in the previous section, penalties associated with scheduling-algorithm, termed BLASOR (Stone et al, 2007).
deviating from production targets introduced in the objective Among other details BLASOR assigns targets in terms of
function aim to control risk of deviation for each element. These produced ore tonnes and grades for each period to each pit as
penalties can be defined in different magnitudes for each element contributing to the global target. Although BLASOR, as used
and period. This enables the mine planner to control the risk for here, accounts for multiple elements, the approach is based on a
each element over time. The ability to control the risk over time is single estimated orebody model and does not incorporate local
a concept introduced by Dimitrakopoulos and Ramazan (2004) uncertainty and in situ variability.
using a geological risk discount rate. This discount rate is directly
applied to penalties and thus controls the risk distribution between Problem specifications and input for scheduling
periods. A high geological discount rate indicates that the SIP The in situ variability and the incomplete knowledge of the
formulation herein is emphasised to generate a schedule that is spatial distribution of the elements in the orebody are most
less risky in early periods than in later periods. This may be useful critical for meeting customer specifications. In order to
when the operation aims to mine less risky parts of the deposits in incorporate in situ variability and uncertainty of geochemical
early periods and more uncertain parts in later periods. As mining parameters in mine production scheduling, techniques for
progresses, more information about those uncertain parts will optimisation under uncertainty can be employed. The application
become available in form of operational exploration. A geological of stochastic mine production scheduling to Yandi Central 1 is
discount rate of zero per cent generates schedules that are based on stochastically simulated orebody models generated
expected to exhibit a similar level of risk in all periods. The using the computationally joint direct block simulation approach
difference between penalties applied to upper deviations and lower (Boucher and Dimitrakopoulos, 2009). Operational, economic
deviations defines the priority of upper and lower deviations from and risk defining parameters are explained in subsequent sections
targets. For example, it may be more important in an operation to in more detail.
keep the deficit in production as low as possible while excess
production may not be of importance.
Stochastic orebody models at Yandi Central 1
PRODUCTION SCHEDULING UNDER The basis for mine production scheduling under geological
UNCERTAINTY – AN APPLICATION AT uncertainty is a series of simulated orebody models of the
YANDI CENTRAL 1 IRON ORE DEPOSIT, deposit. For this case study, 20 simulated orebody models of the
WESTERN AUSTRALIA main ore zone (MOZ) are used, generated by Boucher (2003)
using the joint block simulation technique detailed in Boucher
Next, mine production scheduling under multi-element grade and Dimitrakopoulos (2009). This joint-simulation of the five
uncertainty is applied to the Yandi Central 1 iron ore deposit in considered elements Fe, P, SiO2, Al2O3 and LOI guarantees the
Western Australia. The first part describes the Yandi Central 1 local reproduction of cross-correlation between the elements.
deposit focusing on geology, mining operation and current Note that Fe is strongly correlated with the elements SiO2 and
production scheduling practice. The problem specification and Al2O3. Each of the resulting orebody models contains 3049
-86500 N Lower bench
-86000 N
-85500 N
-13000 E
-15000 E
-11000 E
-86500 N Middle bench
-86000 N
Fe-content in %
-85500 N
-13000 E
-15000 E
-11000 E
0.0-57.4
57.4-58.2
58.2-58.7
58.7-60.6
Upper bench
-86500 N
-86000 N
-85500 N
-13000 E
-15000 E
-11000 E
FIG 2 - Spatial distribution of Fe-grades in realisation number five for the lower, middle and upper bench.
blocks in total. Block dimensions are 25 m × 25 m × 12 m, 0.56 per cent. Further, it is assumed that the operation is flexible
representing typical mining units. Each block contains the enough to account for different ore and waste production rates
attributes total tonnage, ore tonnage as well as total content of between periods. For this reason, the maximum mining capacity,
each element Fe, P, SiO2, Al2O3 and LOI. As an example, a map including ore and waste production, was set to 20 000 000 t,
of the spatial distribution of Fe grades in the orebody model is which is about 5 000 000 t more than the maximum rate. Due to
presented in Figure 2 for the case of simulated realisation the flat geometry of the deposit, one slope region is sufficient to
number five. characterise the geotechnical constraints. The general slope angle
is set at 45°.
Operational parameters
Economic and risk-controlling parameters
Operational parameters, including ore production and required
qualities are defined by the global multi-pit scheduling approach Table 2 presents the economic parameters, including price,
undertaken by BLASOR. BHP Billiton Iron Ore provided mining and processing costs and discount rates. Mining costs
scheduling results defining the contribution of Yandi Central 1 to include blasting, extraction and transportation costs – processing
the global target for the following five years referred to as periods. costs account for crushing, conveying and stockpiling. Two
For confidentiality reasons, BLASOR results are scaled (Table 1).
discount rates are identified, the economical discount rate and
Ideally, shipping grades are to be delivered with nearly zero the geological discount rate. The economical discount rate
variability. Since this is unlikely, the industry sets target bands
limited by an upper and lower bound. Grades should not fall discounts cash flows over periods, while the geological discount
outside this band. Table 1 summarises initial ore tonnage and rate controls the risk of producing grades that fall outside the
grade limits. The differentiation between ore and waste prior to limits over the periods. Recovery is 100 per cent. The stochastic
the optimisation is realised through an Fe grade cut-off of scheduling approach applied in this case study is concerned with
TABLE 1
Ore tonnage and grade constraints for scheduling Yandi Central 1.
BLASOR scheduling results of Yansi Central 1 for first periods

Page number Ore tonnage (wt) Fe (%) P (%) SiO2 (%) Al2O3 (%) LOI (%)
1 14 000 000 57.1 - 59.4 0.032 - 0.038 4.6 - 5.2 0.90 - 1.05 9.5 - 11.0
2 10 000 000 57.1 - 59.4 0.032 - 0.038 4.6 - 5.2 0.90 - 1.05 9.5 - 11.0
3 10 000 000 57.1 - 59.4 0.032 - 0.038 4.6 - 5.2 0.90 - 1.05 9.5 - 11.0
4 9 000 000 57.1 - 59.4 0.032 - 0.038 4.6 - 5.2 0.90 - 1.05 9.5 - 11.0
5 7 200 000 57.1 - 59.4 0.032 - 0.038 4.6 - 5.2 0.90 - 1.05 9.5 - 11.0
Note: Ore/waste cut-off grade is Fe> = 56 per cent.
define the pit design for each period and provide a mineable
TABLE 2 production schedule. Parameters used in this designing process
Economical parameters for long-term production scheduling of the are a 12 m bench height, 45° slope angle and a 5 m berm
Yandi Central 1 iron ore operation.
between two toe and crest string, a road width of 25 m and a
Parameter Cost/price eight per cent ramp incline. The lower part of Figure 3 shows a
south-east isometric view of the resulting smooth schedule.
Price per ton recovered metal $30
Benndorf (2005) demonstrated that this type of smoothing has no
Mining costs per ton $5 significant impact on the results, which means that the smoothed
Processing costs per ton $5 schedule is still near to optimal.
Economical discount rate 10%
Geological discount rate 10% Evaluating results
In addition to produced ore and waste tonnage, results are

the risk of not meeting production targets of produced evaluated in terms of risk profiles of produced grades per period,
element-grades. Penalties for deviating from production targets in particular for Fe, SiO2, Al2O3, P and LOI (Figure 4). For each
are set initially to $1/unit of deviation. period the grades are shown considering each simulated orebody
realisation, which represent possible scenarios based on
THE PRACTICAL SCHEDULING APPROACH information available. The spread of the different realisations
provide an indication about uncertainty in produced grades per
period when extracting the deposit according to the generated
Initial run and practical mine design
schedule. Analysing the risk profiles of Fe, P and LOI results
The upper part of Figure 3 shows results of an initial run using concludes that there is no risk of deviating from production
above specified parameters. The extraction sequence appears targets. SiO2 and Al2O3 appear to be more critical in meeting
smooth and feasible, however there are few blocks scheduled production targets. For example, four out of twenty simulated
surrounded by blocks scheduled in different periods. To generate orebody models for SiO2 indicate a deviation from the lower
a practical mining schedule that guarantees minimum mining target in period one. Thus, there exists a 20 per cent chance of
width and equipment accessibility, results of the stochastic not meeting production targets for SiO2 in period one.
formulation are refined using manual mine design and haul road
construction. These standard tools are available in commonly The ability to control risk
used mine scheduling software packages. In this study open pit
design from Earthworks Datamine is used (Data- mine manual, A major contribution of the presented scheduling formulation is
2002). The schedule generated by the formulation can be used as the ability to control risk of deviating from production targets
a guideline to construct polygons for each period and bench. considering different quality parameters. As experienced in the
These polygons, in combination with haul roads and ramps, initial run, SiO2 and Al2O3 appear most critical in meeting
targets. To investigate the ability to decrease risk, three different
schedules were generated applying different penalties to both
critical elements. The three schedules were generated using low
(1$ per unit deviation per tonne), medium (10$ per unit deviation
per ton) and high penalties (100$ per unit deviation).
Figure 5 shows the extraction sequence of the lower bench for
each schedule. In the case of each schedule, the deposit would be
extracted in a different sequence. The dispersion of the schedules
increases with the magnitude of the penalties. In the case of low
penalties, the extraction sequence is smooth. Although medium
penalties generate a more dispersed schedule, it is still smooth
Period 1
enough to be converted to a feasible schedule using manual mine
Period 2 design. High penalties generate a very dispersed schedule, which
0m Period 3
could hardly be efficiently realised. The dispersion is an
50 Period 4
Period 5 expression of a higher selectivity, necessary in order to produce a
homogeneous product in a tight quality band. Figure 6 shows the
risk profiles for SiO2 and Al2O3 for the three generated schedules.
In case of SiO2, the effect of increasing penalties already becomes
obvious in the case of medium value penalties. Compared to the
low penalty case, the fluctuation of grades between periods
decreases significantly and there exists only a slight probability of
deviating from targets in period 2, 3 and 4. Higher penalties
improve the result only marginally. In the case of Al2O3, a
decrease in probability of deviating from targets is recognisable
with higher penalties, however, there still exists a certain amount
Period 1
Period 2
of risk. This is an expression of a high in situ variability and
0m Period 3 uncertainty of the element, which cannot be avoided by blending
50 Period 4 in the pit. A solution here, to decrease the risk, could be to blend
Period 5
the ore with ore from different mines, where Al2O3 is less variable
and uncertain.
FIG 3 - Stochastic schedule in ultimate pit – before (upper part) Generally, this evaluation of the scheduling formulation
and after (lower part) smoothing using manual design. demonstrates that less risk of deviation comes with a cost of
Production Yandi Central 1 Fe Grades

Limits: 57.1 % - to 59.4 %
16 59.50%
Millions
14
59.00%
12
10 58.50%
Grade in %
Ore
Tons
8 Waste
Total 58.00%
6
4
57.50%
2
0 57.00%
1 2 3 4 5 0 1 2 3 4 5 6
Period Period
Silica Grades Alumina Grades

Limits: 4.6 % to 5.2 % Limits: 0.90 % to 1.05 %
5.3% 1.05%
5.1% 1.00%
Grade in %
Grade in %
4.9% 0.95%
4.7% 0.90%
4.5% 0.85%
0 1 2 3 4 5 6 0 1 2 3 4 5 6
Period Period
LOI P-Content
Limits: 10.0 % to 11.0 % Limits: 0.032 % - to 0.038 %
10.70% 0.038%
0.037%
10.60%
Grade in %
Grade in %
0.036%
10.50%
0.035%
10.40% 0.034%
0.033%
10.30%
0 1 2 3 4 5 6 0.032%
Period 0 1 2 3 4 5 6
Period
FIG 4 - Results of stochastic scheduling in terms of ore and waste tonnages and risk profiles for Fe, SiO2, Al2O3, P and LOI.
Lower bench
Low penalties ( 1 per unit deviation)
Medium penalties (10 per unit deviation)
High penalties (100 per unit deviation)

Period 1
Period 2
Period 3
Period 4
Period 5
FIG 5 - Different extraction schedules depending on the magnitude of penalties for the lower bench.
Alumina Grades
Limits: 0.90 % to 0.95 %
1.05%
Low penalties ( 1 per unit deviation)

1.00%
Grade (in %)
0.95%
Alumina Grades
Limits: 0.90 % to 0.95 %
0.90%
1.05%
0.85%
1.00%
0 1 2 3 4 5 6
Grade (in %)
Period
0.95%
Medium penalties (10 per unit deviation) 0.90%
0.85%
Alumina Grades
0 1 2 3 4 5 6
Limits: 0.90 % to 0.95 %
Period
1.05%
1.00%
Grade (in %)
0.95%
0.90%
0.85% Limits used in SIP formulation

0 1 2 3 4 5 6
Period
Silica Grades
Limits: 4.7 % to 5.1 %
5.3%
Low penalties (1 per unit deviation)

5.1%
Grade (in %)
4.9%
Silica Grades
4.7% Limits: 4.7 % to 5.1 %
5.3%
4.5%
0 1 2 3 4 5 6 5.1%
Period
Grade (in %)
4.9%
Medium penalties (10 per unit deviation) 4.7%
4.5%
Silica Grades 0 1 2 3 4 5 6
Limits: 4.7 % to 5.1 % Period

5.3%
5.1%
Grade (in %)
4.9%

4.7%
4.5% Limits used in SIP formulation

0 1 2 3 4 5 6
Period
FIG 6 - Risk profiles for produced grades (alumina and silica) depending on penalties.
higher selectivity, which is caused by the two competing geological uncertainty, is presented and successfully applied to
objectives in the objective function: minimise risk of deviating production scheduling at the Yandi Central 1 deposit, WA. It is
from production targets, and generate a smooth schedule. demonstrated that the SIP formulation presented, can be
implemented as part of a multi-pit scheduling approach. In this
application, results from BLASOR, a multi-pit scheduling
CONCLUSIONS
optimisation approach, are used to define the contribution of the
A new stochastic integer programming based mine production Yandi Central 1 deposit, Western Australia, to the global target per
scheduling approach, which considers jointly multi-element period in terms of desired grades of elements and ore tonnages.
Results demonstrate the ability of the stochastic approach to Dowd, P A, 1997. Risk in minerals projects: Analysis, perception and
control risk of deviating from production targets for critical management, Transactions of the Institutions of Mining and
quality defining elements. A comparison between the stochastic- Metallurgy, Mining Technology, 106:A9-A18.
ally generated production schedule and a schedule generated Godoy, M C and Dimitrakopoulos, R, 2004. Managing risk and waste
using one estimated orebody model illustrated the benefit, mining in long-term production scheduling, SME Transactions,
stochastic models can generate. The stochastic schedule shows a 316:43-50.
higher probability in meeting production targets, which decreases Goovaerts, P, 1997. Geostatistics for Natural Resources Evaluation,
overall project risk and can increase project value. 483 p (Oxford University Press: New York).
ACKNOWLEDGEMENTS deposit, Transactions of the Institutions of Mining and Metallurgy,
Funding was partially provided by NSERC Discovery Grant
Leite, A and Dimitrakopoulos, R, 2010. Production scheduling under
239019-06 and BHP Billiton, Rio Tinto, AngloGold Ashanti and
metal uncertainty – Application of stochastic mathematical
Xstrata. programming at an open pit copper mine and comparison to
REFERENCES conventional scheduling, in Advances in Orebody Modelling and
Strategic Mine Planning I (ed: R Dimitrakopoulos), pp 35-40 (The
Benndorf, J, 2005. Efficient sequential simulation methods with Australasian Institute of Mining and Metallurgy: Melbourne).
implications on long-term production scheduling, MPhil thesis
Ramazan, S and Dimitrakopoulos, R, 2004. Recent applications of
(unpublished), The University of Queensland, Brisbane, 225 p.
operations research in open pit mining, SME Transactions,
Birge, J R and Louveaux, F, 1997. Introduction to Stochastic Pramming, 316:73-78.
421 p (Springer-Verlag: New York).
Ramazan, S and Dimitrakopoulos, R, 2007a. Production scheduling with
Boucher, A, 2003. Conditional joint simulation of random fields on block uncertain supply: A new solution to the open pit mining problem,
support, MPhil thesis (unpublished), University of Queensland, COSMO research report, number 2, pp 257-294.
Brisbane, 168 p.
Ramazan, S and Dimitrakopoulos, R, 2007b. Stochastic optimisation of
Boucher, A and Dimitrakopoulos, R, 2009. Block-support simulation long-term production scheduling for open pit mines with a new
of multiple correlated variables, Mathematical Geosciences, integer programming formulation, in Orebody Modelling and
41(2)142-158. Strategic Mine Planning, second edition (ed: R Dimitrakopoulos),
Chiles, J P and Delfiner, P, 1999. Geostatistics, Modelling Spatial pp 385-392 (The Australasian Institute of Mining and Metallurgy:
Uncertainty, 695 p (John Wiley and Sons: New York). Melbourne)
Datamine manual, 2002. Datamine-guide graphical mining software Ravenscroft, P J, 1992. Risk analysis for mine scheduling by conditional
reference manual. simulation, Transactions of the Institutions of Mining and
David, M, 1977. Geostatistical Ore Reserve Estimation, 364 p (Elsevier: Metallurgy, Mining Technology, 101:A104-A108.
Amsterdam). Stone, P, Froyland, G, Menabde, M, Law, B, Pasyar, R and Monkhouse,
David, M, 1988. Handbook of Applied Advanced Geostatistical Ore P, 2007. Blasor-blended iron ore mine planning optimisation at
Reserve Estimation, 216 p (Elsevier: Amsterdam). Yandi, in Orebody Modelling and Strategic Mine Planning, second
Dimitrakopoulos, R, 2007. Risk analysis in ore reserves and mine edition (ed: R Dimitrakopoulos), pp 133-136 (The Australasian
planning: Conditional simulation concepts and applications for the Institute of Mining and Metallurgy: Melbourne).
mining industry, The Australasian Institute of Mining and Metallurgy Whittle, G, 2007. Global asset optimisation, in Orebody Modelling and
– McGill 2007 Professional Development Seminar Series, 385 p. Strategic Mine Planning, second edition (ed: R Dimitrakopoulos),
Dimitrakopoulos, R, Farrelly, C and Godoy, M, 2002. Moving forward pp 331-336 (The Australasian Institute of Mining and Metallurgy:
from traditional optimisation: Grade uncertainty and risk effects in Melbourne).
open pit design, Transactions of the Institutions of Mining and
Dimitrakopoulos, R and Ramazan, S, 2004. Uncertainty based production
scheduling in open pit mining, SME Transactions, 316:106-112.
HOME
Overall Optimisation of Yerilla Project

T Dincer1 and K Reynolds2
ABSTRACT offer significant technological and economical advantages in the

Yerilla Project, comprising the development of three relatively large
current project, including: increased leach feed grade after
nickel laterite deposits, is located 130 km northeast of Kalgoorlie. Boyce beneficiation of fines, increased range of ore types amenable to
Creek and Aubils deposits are located 25 km and 55 km respectively to leaching, lower acid and calcrete unit consumptions, lower nickel
the northwest of the project site, centred close to the main Jump-up Dam in circuit inventory, and full scale capacity could be attained with
deposit. Planning the development of multiple deposits added to the minimal ramp-up period.
challenges of a geologically and metallurgically complex project in the The key to the success of the Yerilla scoping study would be the
demanding industrial environment in early to mid 2008. The Yerilla integration of process performance and cost parameters into the
Project scoping study completed in September 2008 has resulted in an
mine optimisation process. The study completed in September
18 years’ production schedule, based on the processing of 3.2 million
tonnes resource per annum to produce average 21 000 tonnes of nickel
2008 (Heron, 2008b) has resulted in an 18 years’ production
and 1400 tonnes of cobalt per annum. The overall optimisation of the schedule based on processing of 3.2 million tonnes resource per
project has been crucial for making decisions during the study and is annum to produce average 21 000 tonnes of nickel and
based on the following interdependent technical studies: optimisation of 1400 tonnes of cobalt per annum. Conventional open pit mining
direct feed and beneficiation parameters, optimisation of cut-off grades method was selected for the development of the open pits based on
and throughput rates, mine sequencing and detailed production a major equipment feet of 180 t excavators and 90 t rear dump
scheduling, and analysis of mining equipment options, capital and trucks. Development of multiple deposits in parallel was required
operating costs. in the progress of the production schedule to supply a process feed
This paper presents the technical analysis approach taken in the Yerilla blend with consistent characteristics, which added to the
Project scoping study and the major findings. It has been demonstrated challenges of a geologically and metallurgically complex project.
that the balance between the interdependent project parameters, hence the The overall optimisation of the project was achieved to a
overall optimisation of the project, have been achieved at a sufficient sufficient level for a scoping study on the basis of the following
level through relatively basic analyses undertaken at the pit optimisation
interdependent technical studies: definition of beneficiation and
and mine sequencing stages of the project. The scheduling of equipment
capital and operating resources, depending on the economic environment atmospheric leach processing parameters, optimisation of cut-off
at the time of the study, would be of particular interest. and cut-over grades and throughput rates, mine sequencing of
multiple deposits and detailed production scheduling, and analysis
of mining equipment options, preliminary capital and operating
INTRODUCTION cost estimates.
Yerilla Project, comprising the development of three relatively
large nickel laterite deposits, is located 130 km northeast of
Kalgoorlie. Boyce Creek and Aubils deposits are located 25 km
and 55 km respectively to the northwest of the project site, centred
close to the main Jump-up Dam deposit (Figure 1). The project is
100 per cent owned by Heron Resources Limited (Heron),
involving the development of multiple open pit mines, a treatment
plant and associated utilities, services and infrastructure to recover
nickel and cobalt into a mixed sulfide product, using BHP Billiton
patented atmospheric leach technology (BHPB Patent
AU2007904228). Nickel-laterite mineralisation at the Jump-up
Dam, Boyce Creek and Aubils deposits is developed over
ultramafic dominated lithologies where basement rocks have been
deeply weathered to form a thick regolith profile, the character-
istics of which are summarised below in Figure 2. Most of the
high-grade nickel mineralisation is developed over serpentinised
ortho/ mesocumulate rocks in the form of goethite, smectite and
chlorite rich clays, while lower grade highly siliceous material is
common over adcumulate rocks. The thicknesses of the regolith
profile components and the mineralisation vary widely depending
on the underlying lithology, the presence of faulting and fracturing
in the bedrock, and the amount of surface stripping by erosion.
The initial Yerilla Project study completed in March 2008
(Heron, 2008a) concluded that while heap leaching was
technically feasible at that level of assessment, it carried high
technical risk and was not economically viable. The atmospheric
leaching in combination with beneficiation of fines was seen to
1. MAusIMM, Principal Mining Engineer, Mining Solutions

Consultancy Pty Ltd, 5 Salter Point Parade, Salter Point WA 6152.
2. MAusIMM, Project Manager, Heron Resources Limited, Level 1, 37
Ord Street, West Perth WA 6005.
Email: [email protected] FIG 1 - Yerilla Project location.
T DINCER and K REYNOLDS
Regolith & Character Over Print Ni % Co % FeO % MgO %
Alluvial Gravels Alluvial cover
Iron/Calcrete/ Calcrete/
Pisolitic Duricrust <0.5 <1.0 <0.05 >40 <20 0.5 to 5
Siliceous Cap Siliceous
Ferrugenous:
Smectite:
Ferrugenous/ varying ratios of 0.3 to 0.5 to 0.05 to 25 to 10 to
Nontronite and 0.5 to 35
Smectite Goethite, Haematite & Massive 1.5 2.5 0.5 60 35
Montmorillonite Massive to
Kaolinite to
Dissem.
Dissem.
Magnesite
Free
& Dolomite
Varying ratios of Smectite, Silica 0.03 to
Saprolite 0.5 to 4 20 to 40 5 to 35
Chlorite/Garnierite and Serpentine 0.5
Saprock Variabily oxidised Serpentine 0.3 to 0.7 0.02 15 to 20 30
Fresh Ultramafic Olivine/Serpentine 0.3 0.01 5 to15 30 to 45
FIG 2 - Generalised regolith profile and mineralisation characteristics.
With the exceptions of Whittle Software used in pit limit beneficiated are processed via a direct leach grinding circuit ahead
optimisations and MineSight, used as the host planning software, of leaching. The atmospheric leach process assessed is a two stage
all the other analysis and scheduling tasks in the study were leach process, as patented by BHP Billiton (Patent: AU2007-
completed through advanced spreadsheets, custom programs and 904228). In the first stage (primary) leach, ferruginous (or limon-
a combination of standard procedures. Almost all the information itic) ore types are leached under conditions of high acid addition,
in the project has been linked electronically through procedures, in agitated tanks at atmospheric pressure and approximately
spreadsheets and basic data files, with references to the time and 100°C. The residence time is relatively short at nominally three
team members involved in the project. The interdependent
hours. In the second stage (secondary) leach, saprolite ore types
project tasks as part of this approach were as follows: test work
are combined with the primary leach residue and leached for a
program to obtain beneficiation performance data, analysis of the
beneficiation performance data and definition of processing further 12 hours (nominal) at approximately 95°C under low acid
parameters, economic analysis of the processing methods and strengths. Under these conditions, dissolved iron from the primary
resource utilisation, analysis of the cut-off grades and throughput leach precipitates as jarosite, generating additional acid which is
rates, mine production and processing schedules and mining utilised by the saprolite leaching reactions.
equipment selection, capital and operating cost estimates.
Beneficiation parameters
With the advancement of the computer technology, most of the
mine design and planning tasks, and consequently the decision- A program of test work was undertaken to define the beneficiation
making process in the mining projects, are largely dependent on performance of the Yerilla ore types. Beneficiation test work,
the specialised technical software chosen for the project. The incorporating scrubbing and attritioning, has been undertaken on a
progress of the project commonly depends on the capabilities total of 34 samples. All major Jump-up Dam ore types and four
and limitations of the specialised software and the experience of Boyce Creek samples were tested. As a result of the beneficiation
the user of the software, especially in the production scheduling testwork programme, the selected beneficiation cut size and
area. Contrary to this, the sections below demonstrate how the performance parameters by ore type are presented in Table 1.
There are three main classifications: ferruginous, transitional, and
overall optimisation of the Yerilla project has been achieved
saprolite.
through a system of basic software tools that facilitated systemic,
relativity and sensitivity analyses. The Yerilla project was not For the Yerilla Project, both ferruginous and transitional ore
dependent on the capabilities of any specialised software at any types feed the primary leach circuit. A simplified flowsheet is
stage. The majority of the analyses for the optimisation of the presented in Figure 3. Figure 4 summarises the beneficiation
project parameters took place in the mine sequencing stage, performance results, including the average upgrade ratio verses
recovery performance for each ore type. The ore types with the
which is spatially and directly linked to the resource model,
flat curves in the lower part of the graph are those which do not
mining areas and dependency between the mining areas (panels).
show good beneficiation performance, being ferrougenous (F),
saprolite-nontronite/chlorite (S_NC) and saprolite-serpentinite
PROCESS PARAMETERS (S_S). These ore types were assigned for direct leach processing
without beneficiation. The ore types with the higher sloped
Beneficiation and leaching flow sheet curves beneficiate well, including ferrougenous-nontronite/quartz
(F_NQ), Transitional samples (T), and saprolite-sepentinite/quartz
(S_SQ). The S_SQ sample set was found to have a bi-modal
Coarse low-grade gangue minerals (typically silica and magnesite distribution of good and poor performing samples so that the
minerals) may be rejected from the ROM ore using relatively average performance level was taken for this group.
simple scrubbing, attritioning and size separation techniques,
leading to an upgraded leach feed fraction. However, for the
Yerilla Project, not all ore types respond positively. In addition, as
Atmospheric leach parameters
head grade increases, increased nickel losses to the reject stream For the optimisation work, leach performance was estimated
can make beneficiation suboptimal. Those ore types not from preliminary data available from earlier studies and from
OVERALL OPTIMISATION OF YERILLA PROJECT
TABLE 1
Ore types and beneficiation performance.
Ore type Description Circuit Cut size Recovery (%)

μm
Mass Ni Co MgO Fe SiO2 AI2O3 CaO Mn
Ferruginous (containing goethite)
F Geothite + kaolinite +/- MnO Direct leach Grind 100 100 100 100 100 100 100 100 100
F_Q + Quartz Scrub & attrit 180 33.4 55.5 52.9 49.3 47.8 25.2 51.3 15.4 42.1
F_NQ + Nontronite or quartz Scrub & attrit 180 55.8 75.3 61.7 67.9 72.9 40.1 76.2 52.2 62.3
F3 + Carbonate Scrub & attrit 180 35.7 46.0 40.2 17.5 41.2 57.5 41.9 12.1 37.8
F_Q3 + Quartz + carbonate Scrub & attrit 180 33.4 55.5 52.9 49.3 47.8 25.2 51.3 15.4 42.1
F_NQ3 + Nontronite or quartz + carbonate Scrub & attrit 180 55.8 75.3 61.7 67.9 72.9 40.1 76.2 52.2 62.3
Transitional (containing geothite + serpentinite)
T_NCS + Chlorite + nontronite Scrub & attrit 180 39.0 61.6 62.6 41.1 58.2 37.3 63.0 47.9 48.8
T_NCSQ + Quartz + chlorite + nontronite Scrub & attrit 180 39.0 61.6 62.6 41.1 58.2 37.3 63.0 47.9 48.8
T_NCS3 + Chlorite + nontronite + carbonate Scrub & attrit 180 39.0 61.6 62.6 41.1 58.2 37.3 63.0 47.9 48.8
T_NCSQ3 + Quartz + chlorite + nont +carb Scrub & attrit 180 39.0 61.6 62.6 41.1 58.2 37.3 63.0 47.9 48.8
Saprolite (containing serpentinite)
S_S Serpentinite Direct leach Grind 100 100 100 100 100 100 100 100 100
S_S3 + Carbonate Scrub & attrit 500 66.8 75.8 81.4 70.6 78.9 56.2 72.0 71.0 81.6
S_SQ + Quartz Scrub & attrit 500 48.5 64.1 65.2 57.1 63.7 41.9 64.8 44.6 62.6
S_SQ3 + Quartz + carbonate Scrub & attrit 500 48.5 64.1 65.2 57.1 63.7 41.9 64.8 44.6 62.6
S_NC + Nontronite/chlorite Direct leach Grind 100 100 100 100 100 100 100 100 100
S_NC3 + Nontronite/chlorite + carbonate Scrub & attrit 500 83.2 91.8 93.7 83.0 87.4 81.2 85.6 48.3 87.7
Acid plant
Mining
ROM Stockpile
Crusher
Conveyor Benefication
Agitated Leaching Tanks
Crushed ore stockpile

waste
Fine ore after crushing
Limestone
Stage 1
Neutralisation
Recycling of process water
Precipitation
Stage
Filtering and drying
Acid addition
Intermediate Product
Bulk Bags
Zero discharge site
Shipping from Esperance
FIG 3 - Process flow sheet.
Average Ore Type Performance for Ni

Scrubbed and Attritioned Unit Operating Profit vs Feed Grade - S_SQ Ore Type
75, 180, 500, 1800, 3350 um cut size 100
1.80
90
80
1.60
70
60
Operating Profit ($/t ROM)

Upgrade Ratio
50
1.40 40 Beneficiation Feed
30 Direct Leach Feed
20
1.20 10
0
-10 0.0 0.2 0.4 0.6 0.8 1.0 1.2 1.4 1.6 1.8 2.0
1.00
-20
20.0 30.0 40.0 50.0 60.0 70.0 80.0 90.0 100.0
Distribution (%Ni) -30
F Scrub F_NQ Scrub T_all Scrub S_SQ good Scrub S_S Scrub -40
S_SQ bad Scrub S_NC Scrub F3 Scrub S-SQ3 Scrub F_NQ Attrit -50
F Attrit F3 Attrit T_all Attrit S_S Attrit S_NC Attrit
Feed Grade (%NiEq)
S_SQ good Attrit S_SQ bad Attrit S-SQ(3) all Scrub S_SQ(3) all Attrit S_SQ3 Attrit
FIG 4 - Beneficiation performance summary. FIG 5 - Cut-off and cut-over grade determination.
knowledge of typical performance from other projects. Leach value compared to the direct leach option. Based on Figure 5, the
extraction parameters used in the optimisation work are cut-off and cut-over grades for the S_SQ ore type are as follows:
presented in Table 2. Final leach test work data was not available 0.55 per cent Ni break-even cut-off grade for the beneficiation
until late in the study schedule.
ROM feed, and 1.10 per centNi cut-over grade for the direct
leach ROM feed.
TABLE 2 The same methodology was applied to determine cut-off and
Leach performance assumptions by ore type. cut-over grades for the other ore types stored in the resource
model. The break-even cut-off grades generally varied between
Ore type Circuit Recovery (%) 0.40 per cent Ni and 0.70 per cent Ni, depending on the process
Ni Co recoveries for the ore types. In order to simplify ore classification
Ferruginous Primary 95 90
criteria, the cut-off and cut-over grades were standardised across
the ore types after checking the economic sensitivities.
Transitional Primary 95 90
S_S and S_S3 Secondary 75 60 Cut-off grade analysis
S_SQ and S_SQ2 Secondary 85 70
Prior to the determination of the ultimate pit limits and
S_NC and S_NC3 Secondary 85 70 production scheduling, a cut-off grade and throughput optimis-
ation study was completed as part of the pit limit optimisations to
determine the optimal production parameters for the project. The
CUT-OFF GRADE AND THROUGHPUT RATE four cut-off grade cases in the study included the break-even
(BE), 0.6 per cent Ni, 0.7 per cent Ni and 0.8 per cent Ni cut-off
Break-even cut-off and cut-over grades grades. Preliminary production schedules were completed for each
cut-off grade case based on the optimal pit shell sequences
Subsequent to the determination of the metallurgical parameters combined from Jump-up Dam and Boyce Creek models available
for the ore types, economical comparison of the beneficiation plus at the time of the analysis. The combinations of 2.0 Mt/a, 2.5 Mt/a
leach and direct leach processing methods was necessary for and 3.0 Mt/a leach plant throughput rates were studied for each
further classification of the ore types for mine planning purposes. cut-off grade case.
This required the calculation of the cut-off and cut-over grades The initial schedules accommodated the variable cut-off grades
depending on the operating costs and recoveries for each process- through time as the cut-off grades for each increment in the
ing method. The cobalt credit for the ore types was included in the mining sequence was selected to provide maximum Net Present
calculations. Leach plant capital reduction benefit for the Value (NPV). However, the project NPV was not significantly
better than the constant cut-off grade cases with relatively higher
beneficiation process was included in the comparison of the
NPVs. In fact, the overall differences between the studied constant
economic values for the calculation of the cut-over grades between cut-off grade cases were not so large (only a few percentage
the beneficiation plus leach and direct leach circuits. The capital points) due to the beneficiation of the low-grade ore within a
reduction benefit was based on the fact that a lower leach plant limited grade range below the cut-over grade. Also considering the
capacity would be required to produce the same amount of metal practicality aspects of the project associated with the number of
as a result of the upgrade through the beneficiation process. The ore types and relatively lower confidence in the early stage of
effect of beneficiation is estimated as a 12.5 per cent overall resource definition, especially for Boyce Creek and Aubils
reduction in the processing plant and related infrastructure capital deposits, a constant cut-off grade strategy was opted in the
costs. Figure 5 shows the variation of the unit operating profit as production schedules.
the feed grade varies for two processing methods. The curves are The graph in Figure 6 shows the variation of the cumulative
for the S_SQ ore type with the most significant tonnage (22 Mt). operating NPV at Years 4, 8 and 12 of the preliminary production
The graph shows that the beneficiation adds value relative to direct schedules for 2.5 Mt/a throughput rate. The analysis of the cut-off
leach option for the low-grade ore below ~1.00 per cent Ni by grades across the throughput rates can be summarised as follows:
upgrading the metal content. Above 1.00 per cent Ni feed grade, Compared to the break-even cut-off, the NPV improvement in the
direct leaching of the ROM ore is better since the metal losses 0.6 per cent Ni cut-off cases is generally not significant, indicating
become more prominent and start reducing the unit operating that some marginally economic resource is still included in the
2500ktpa Leach Cumulative Operating NPV improvements in the operating NPV as the throughput rate
140% increases, the capital cost of the processing plant capacity
0.8%Ni COG
increase needs to be taken into account. The same curves in
135%
0.7%Ni COG Figure 8 are based on the NPV values normalised for the plant
130% 0.6%Ni COG and infrastructure incremental capital costs. The throughput rate
B.E. COG
125% analysis for the normalised NPVs showed that the overall NPV
Operating NPV (%)
120%
of the project would not be improved significantly above a
2.5 Mt/a throughput rate. The NPV for the 3.0 Mt/a throughput
115%
rate is significantly better at Year 8, but drops to a lower level
110% than the 2.5 Mt/a case at Year 12 of the schedule. The shorter
105% mine life for 3 Mt/a case would also affect the depreciation of the
100%
plant and infrastructure capital, slightly increasing the overall
(Year 12) NPV gap between 2.5 Mt/a and 3.0 Mt/a cases. Subject
95%
to further review in future, the throughput analysis results
90% showed that a 2.5 Mt/a throughput rate would satisfy the
3 4 5 6 7 8 9 10 11 12 13
production requirements for the project and improve the project
Years
NPV slightly in the first 12 years of operations.
FIG 6 - Operating NPV for cut-off grade cases. 0.7%Ni Cut-off Normalised Cumulative Operating NPV
140%
3000ktpa Leach
schedules; Compared to the 0.6 per cent Ni cut-off, the 0.7 per 135% 2500ktpa Leach
2000ktpa Leach
cent Ni cut-off grade provides significant improvement in the 130%
operating NPV values, eight per cent in Year 8 and four per cent
125%
in Year 12; and Although the 0.8 per cent Ni cut-off grade
Operating NPV (%)

improves the NPV more significantly at the end of Years 4 and 8 120%
(12 per cent and 16 per cent respectively), the overall NPV is 115%
lower than the 0.7 per cent Ni cut-off case due to the
110%
under-utilisation of the available resources. The shorter mine life
would also reduce the depreciation time for the plant and other 105%
project infrastructure. 100%
The low-grade resources below cut-off grade were stockpiled 95%

but not considered for treatment within the scope of the project
90%
due to the marginal economics, long mine life (close to 20 years) 3 4 5 6 7 8 9 10 11 12 13
and the insignificant impact on the project NPV. The processing Years
of the low-grade resources would require the feasibility of a new
follow-up project with the increased beneficiation and leaching FIG 8 - Normalised NPV for throughput rate cases.
capacity to achieve sufficient levels of metal production rates.
Further potential resource definition in the project area, as well
as likely changes in the mineral economics in the 20+ years time Cut-off grade and throughput optimisation
frame, did not justify the inclusion of the low-grade treatment results
within the scope of the current project.
The cut-off grade and throughput rates analyses discussed in the
Throughput rate analysis two subsections above are related and all the combinations of the
cut-off and throughput cases were studied. The following is a
For the selected 0.7 per cent Ni cut-off grade, the variation of the summary of the combined findings from the review of cut-off
cumulative operating NPV at the end of the Years 4, 8 and 12 of grade and throughput rate production analysis results:
the production schedules can be seen in Figure 7 for the leach
capacity cases studied. Although the curves show significant • the proposed 0.7 per cent Ni cut-off grade removes the
marginally economical resources from process feed and
brings some cash flow earlier without compromising the
0.7%Ni Cut-off Cumulative Operating NPV overall value of the project;
140% • the cut-off grades above 0.7 per cent Ni reduce the proportion
3000ktpa Leach
135% 2500ktpa Leach
of the lower grade beneficiation feed significantly, hence
2000ktpa Leach reducing the overall utilisation of the low-grade resources;
130%
125%
• the beneficiation of the low-grade ore does not improve the
operating NPV of the project, but helps to reduce the leach
Operating NPV (%)
120%
plant capacity (hence lower capital) to achieve the same
115% nickel production;
110% • the review of the resource models have shown that the
105% resource outlines become less continuous as the cut-off grade
100%
increases above 0.6 per cent Ni, which would in turn increase
the low-grade dilution, or require more selective mining; and
95%
90%
• the 2.5 Mt/a total leach rate case produces higher, or at least
3 4 5 6 7 8 9 10 11 12 13 similar overall NPV (including capital), compared to the
Years 2.0 Mt/a and 3.0 Mt/a cases.
As a project decision based on the findings above, the
FIG 7 - Operating NPV for throughput rate cases. combination of the 0.70 per cent Ni cut-off grade and 2500 kt/a
total leach feed rate provides the best value for the project. The
beneficiation rate, secondary/primary leach rates and annual
nickel production for this combination, calculated as part of the Jump-up Dam
analyses, are as follows: 1.4 Mt/a of average beneficiation rate
providing 0.7 Mt/a of beneficiated leach feed rate, 1.8 Mt/a of
average direct leach feed rate for a total of 2500 kt/a total leach
rate, and 22 000 t/a average nickel production in the first
13 years, then 20 000 tpa average for the rest of the mine life.
Note that the Aubils resource model was not ready and not
included in the study at the time of the production analysis. It is
currently planned to develop the Aubils deposit after eight years
of plant operation, so the results from the production analysis
above will be critical for cash flow in the early years. Reviews of
the production schedule and Aubils open pit economics indicated
that the inclusion of the Aubils deposit is likely to strengthen the
case for the selected cut-off grade and throughput rate.
PRODUCTION SCHEDULE
Boyce Creek Aubils
Mine sequencing and scheduling
Detailed mine designs were not justified at this stage of the study
due to the generally shallow nature (40 - 80 m depth) of the pit
shells. The pit slopes allowed for the ramps and the smoothing
effect in the pit design process would not be significant on the
reported quantities, considering the relatively large extend of the
resource areas. Although the Aubils pit quantities, due to the
discontinuous nature of the open pit limits, may be affected
significantly as a result the pit design process, this deposit was
scheduled late in the mine life and the resource definition was in
an early stage at the time of the study. The optimal pit areas for
each deposit have been divided into the production panels (as
development stages) that can be developed relatively indepen-
dently. The internal variation in the optimal pit shell surfaces
(highs and lows) were taken into account in the definition of the
panel boundaries to form a practical mining sequence. The FIG 9 - Operating NPV for cut-off grade cases.
economically more significant (low cost or higher value) areas
were delineated within the panel boundaries, depending on the pit throughout the mine life, as well as to provide a basis for the
optimisation results. The resultant production panels for the long-term cash flow analysis. The resultant long-term production
Jump-up Dam, Boyce Creek and Aubils deposits can be seen in schedule is relatively flexible, and the short-medium term
Figure 9. Each panel represents three to six months of mine production tonnage and grade variations around the long-term
production in terms of tonnages, but can be developed within six schedule are expected to lie within an acceptable range.
to 12 months due to the mining of multiple panels for blending
purposes and limitations with the vertical advance rates. Production schedule results
The production scheduling process allowed mining of the
multiple panels from multiple deposits to achieve the schedule The total tonnes and ROM ore tonnes scheduled by deposits can
objectives. An initial mining sequence was defined at the start of be seen in Figure 10. The annual production rate increases
the scheduling process based on ranking the economics of the gradually in the schedule as the areas with higher stripping ratios
production panels. The initial sequence providing the best developed towards the end of the mine life. The variation in
operating NPV was modified in the scheduling process, mining rates was smoothed, but is still largely dependent on the
depending on the priority of the objectives and constraints. processing requirements rather then the maximum utilisation of
Consistent primary and secondary leach feed ratios, consistent the mining equipment (discussed in mining equipment section
utilisation of the beneficiation plant capacity, variation of grades below). The ROM ore production in the schedule varies between
(Ni, Co, CaO, Mn, FeO, SiO2, MgO, CO3, Al2O3) within a 3.0 Mt/a and 3.5 Mt/a, depending on the beneficiation feed rates,
reasonable range, deferment of the mining costs and earlier higher mining rates and variation in the stockpile balances. Figure 11
cash flow were the main objectives of the production schedule. shows that the beneficiation rates vary between 1.2 - 1.7 Mt/a
The production scheduling methodology comprised four phases in and the direct leach feed rates vary between 1.7 - 2.0 Mt/a. The
order to balance the mine, mill and stockpiling schedules as beneficiation feed ROM grade averages at 0.81 per cent Ni
follows: an initial ore supply schedule was completed to balance through mine life and the average grade for direct leach ROM
the production of the ore types to achieve relatively smooth ore is 1.05 per cent Ni. The beneficiated leach feed grade is
primary to secondary leach ratios around 0.9, the processing slightly higher than the direct leach feed grade after 30 - 35 per
schedule was adjusted to achieve relatively constant beneficiation cent upgrade in the beneficiation process. Beneficiated leach
to direct leach ROM feed ratios through project life, the feed constitute 27 per cent of the total leach feed in the schedule,
processing and stockpiling schedules were finalised within the providing 20 per cent reduction in the leach plant capacity and
grade range constraints in the third phase to achieve relatively related operating cost savings in low-grade ore processing.
smooth (decreasing in stages) nickel production over mine life, As seen in the last graph in Figure 11, the schedule achieves
and in the final phase, the production from different panels and relatively constant primary and secondary leach feed tonnages
deposits was balanced to regularise the mining rates with the and grades throughout mine life. The primary and secondary
working stripping ratio increasing gradually over mine life. leach feed grades will be similar in the first seven years when the
The details provided in the production schedule are considered relatively higher-grade areas of the Jump-up Dam and Boyce
sufficient to highlight any potential operational difficulties Creek deposits will be mined. The secondary leach feed grades
Total Tonnes Mined by Deposit Beneficiation and Direct Leach ROM Ore Feed
4,000 1.60
22,500
20,000 3,500 1.40
17,500 3,000 1.20
G rade (N i% )
Tonnes'000
15,000 2,500 1.00
Tonnes’ 000
12,500 2,000 0.80

10,000 1,500 0.60
7,500
1,000 0.40
5,000
500 0.20
2,500
0 0.00
0
-1 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18
-1 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 Years
Years
Direct Leach ROM Feed Beneficiation Feed ROM Feed
Jump-up Dam Boyce Creek Aubils Direct Leach ROM Grade Beneficiation ROM Grade
Beneficiated Leach and Direct Leach Feed

ROM Ore Tonnes Mined by Deposit 3,000 1.50
4,000
2,500 1.25
3,500
3,000 2,000 1.00
Grade (Ni% )
Tonnes'000
2,500
Tonnes’ 000
1,500 0.75
2,000
1,000 0.50
1,500
1,000 500 0.25
500
0 0.00
0 -1 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18
-1 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 Years
Years Direct Leach ROM Feed Beneficiated Leach Feed
Direct Leach ROM Grade Beneficiated Leach Feed Grade
Jump-up Dam Boyce Creek Aubils
Primary and Secondary Leach Feed
3,000 1.50
FIG 10 - Mine production schedule.
2,500 1.25
2,000 1.00
G rade (N i% )
Tonnes'000
will be slightly lower after Year 7, when the remaining resource

1,500 0.75
with relatively low-grade high saprolite proportion is blended in
to the Aubils resource with high primary feed proportion. The 1,000 0.50
annual nickel production in the schedule generally varies 500 0.25

between 22 000 and 24 000 tonnes (Figure 12). The average
0 0.00
cobalt production is 1400 tonnes per annum, raising the total -1 1 2 3 4 5 6 7 8
9 10 11 12 13 14 15 16 17 18
metal production close to 25 000 tonnes per annum except the Years
last four years. The total metal production gradually drops to Primary Leach Feed
Primary Leach Feed Grade
Secondary Leach Feed
Secondary Leach Feed Grade
21 000 tonnes of metal in the last four years when the low-grade
and high stripping ratio areas of the three deposits will be mined.
Note that these metal production values are based on the FIG 11 - Beneficiation and leach feed schedules.
preliminary leach recovery estimates, which have been reduced
by approximately ten per cent in the final estimates used in the Annual Nickel and Cobalt Production
project’s financial analysis. 30.0
25.0
MINE EQUIPMENT AND COSTS ANALYSIS
20.0
Tonnes’ 000
The major mining equipment selected for the project comprises

15.0
180 tonne excavators and 90 tonne capacity rear dump trucks. The
excavator dipper capacity was selected ten per cent larger than the 10.0
standard capacity, depending on the lighter density of the laterite 5.0

and other oxide material. The variation of the excavator and truck
0.0
numbers through mine life are shown in Figure 12. The mining -1 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18
fleet numbers increase following mainly the increase in the Years
stripping ratio and partly in the number of deposits mined Nickel Ni Eq. Cobalt
simultaneously. Considering the number of deposits and panels to
be mined in parallel, the equipment fleet has been standardised
FIG 12 - Nickel and cobalt production schedule.
despite the uneven increase in the equipment requirements
throughout mine life. One of the excavators was scheduled for a
single shift per day in some years and the proposed equipment reclamation from the relatively large stockpiles may be necessary
numbers allow forward purchase of the equipment to provide to regulate the processing feed quantity and qualities (especially at
additional flexibility in the production rates. Boyce Creek and Aubils, where the ore needs to be rehandled
As shown in Figure 12, the excavator and truck fleet utilisation anyway). Contrary to the approach discussed above, in a tight
in the schedule provides 20 - 30 per cent extra capacity through equipment capacity case, the nature of the laterite operations is
mine life for flexibility and effective on-the-job utilisation. The likely to require the mobilisation and hire of additional equip-
waste and/or ore mining campaigns will/may also demand ment to overcome production difficulties. A preliminary cost
additional equipment capacity up to several months during comparison indicated that the capital cost for the additional
operations. Although the available equipment capacity should equipment capacity would be justified especially for effective (on
reduce the stockpiling requirements, the construction and the job) utilisation of the equipment, saving fuel and labour costs
Although the approach above is sometimes called

No of Excavators, Left Axis
No of Trucks, Left Axis
‘conventional’, ‘old fashioned’ or ‘manual’ by advocates of
20 100%
Scheduled Time Utilisation, Right Axis specialised mining software, the system of procedures, data and
18 90%
Available Time Utilisation, Right Axis
analysis tools can be linked and executed to reach a project
16 80%
balance efficiently. Contrary to this approach, it is common that
14 70%
the planning of the laterite (and other complex) deposits is
No of equipment
Utilisation (%)
12 60%
generally limited with the capabilities of the available
10 50%
optimisation/scheduling software and experience of their users.
8 40%
The first objective of project development, especially at the
6 30%
scoping stage, should be to achieve an overall balance between the
4 20%
project parameters. Such a project balance, with associated
2 10%
limitations, would also provide scope for the employment of any
0 0%
-1 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18
specialised software for further project development. Considering
Years the usual limitations with the available data at the project stage,
and future economic uncertainties through the life of the project,
FIG 13 - Mining equipment requirements. being ‘approximately right practically’ in an overall project sense
should be much better compared to being ‘absolutely right
theoretically’ in a specific area of the project. It is also suggested
through mine life. Reducing the non-productive operating time of
that results from optimisation studies, especially production
the mining equipment would also have the benefit of reducing the
schedule related studies, should be compared to a practical base
ancillary equipment utilisation and areas required for large
case to acknowledge the contribution of the software and justify
stockpiles. The NPV of the capital cost of purchasing mining
any changes, if any, for the increased complexity. The Yerilla
equipment earlier in the schedule and under-utilisation was
Project experience and the results obtained have also again shown
compared to the NPV of the potential rehandling costs. This
(after Dincer and Peters, 2001; and Dincer, 2001) the importance
analysis indicated that the NPV of the additional equipment
of the basic pit limit optimisation and mine sequencing processes
capacity at all times is approximately equal to the NPV of the
in mine planning. Considering the interdependency of the project
rehandling of 20 per cent of the ROM ore. This would drop to less
parameters and iterative analyses required to reach a balance,
than ten per cent rehandling of the ROM ore when the additional
these two basic processes can effectively: link geological,
capital cost of the stockpile rehandling equipment (eg an extra
metallurgical, mining and economical parameters both
loader and two trucks) was considered in the analysis.
structurally and procedurally; provide a range of physical and
economical quantities with spatial/geometrical references for the
CONCLUSIONS whole resource area; facilitate a wide range of evaluations and
sensitivity analyses on the project parameters; and potentially
The economic analysis of the Yerilla Project indicated that an provide foundations for an optimised mining project and
increase in the capital costs for a reduction in the operating costs production schedule.
would improve the value of the project – simultaneously resulting
in the planning of more environmentally responsible mining ACKNOWLEDGEMENTS
operations. This result is not surprising considering the economic
conditions and trends in other industries at the time of the study. The authors would like to thank Heron Resource Limited for
The scoping study results do not claim to solve all the issues, permission to publish this paper. The authors would also like to
especially related to the variability in the project. The results are acknowledge Mr James Ridley for his contribution to the project
mostly limited with the available information, which needs to be in the areas of geology and resources.
reviewed, revised and further detailed as the confidence on the
resources and other project parameters increase. REFERENCES
The approach taken and the results obtained in the Yerilla BHP Billiton, 2007. Atmospheric acid leach process for laterites
project have demonstrated that overall balance between the Australian patent: AU2007904228.
geological, metallurgical and mining parameters in a project can Dincer, T, 2001. Application of pit optimisation algorithms beyond open
be achieved through basic analyses linked spatially to the resource pit limits, in Proceedings 17th International Mining Congress and
and mining areas. The overall optimisation of the Yerilla Project Exhibition of Turkey, pp 549-556.
was effectively achieved in the scoping study through: Dincer, T and Peters B, 2001. Blending optimum pit mining sequences,
understanding the spatial geological and metallurgical variability, in Proceedings Strategic Planning Conference, pp 43 - 54 (The
systematic analysis of the cases/options, task or problem Australasian Institute of Mining and Metallurgy: Melbourne).
isolation, relative analyses and comparisons, sensitivity analysis Heron Resources Ltd, March 2008. Jump-up Dam project prefeasibility
on the findings to further test the variability, and concluding on the study, internal report.
trends and ranges with known limitations rather than specific Heron Resources Ltd, September 2008. Yerilla project scoping study,
internal report.
numbers.
HOME
Investigating the Variability of Key Coal Quality Parameters in

Continuous Mining Operations when Using Stockpiles
J Benndorf1
ABSTRACT
1. cut
The delivery of in-spec coal qualities is essential for an efficient and
environmental friendly operation of modern coal-fired power plants. The
design of the mining operation and blending opportunities plays a key
role in homogenising variability and improving the prediction of key 2. cut
quality parameters, such as the calorific value (CV). The presented case
study uses the technique of conditional simulation in geostatistics to
Coal stock and

investigate the variability of the CV and its behaviour along the
blnding yard
extraction, transportation and blending process in a continuous mining
environment. Results of the study enable the opportunity to optimise the
operation mode and adjust it according to the customer’s requirements.
INTRODUCTION
The efficient and environmental friendly operation of modern
Train load
power plant technologies calls for strictly meeting the customer’s
specification of key coal quality parameters while maintaining a
low variability of these parameters. The daily mine production FIG 1 - The continuous mining system under study.
aims to meet the customer specifications in terms of tonnage and
coal quality parameters for different products while minimising
specific costs. The key objective of complex continuous mining In total, the mining system can be operated in four operating
systems is to achieve in-spec deliveries at the lowest cost, by modes:
optimal blending from different faces and multiple seams with a 1. operating only the first cut with a bucket wheel excavator
diverse quality distribution. The blending process typically without utilising the blending yard,
involves the operational mode of the excavators, the conveying
system, as well as stacking and reclaiming at stock and blending 2. operating only the second cut with a chain excavator
yards. The application of conditional simulation in coal mining without utilising the blending yard,
operations is not very well documented, with the exception of 3. operating both cuts at the same time without utilising the
some simplified case studies (Hohn and McDowel, 2001; Costa, blending yard, and
Zingano and Koppe, 2000; others). The technique of conditional
simulation (Goovaerts, 1997) with its ability to map in-place 4. operating both cuts at the same time and utilising the
variability and uncertainty in the prediction of coal quality blending yard.
parameters is suited to investigate the ability of the The study investigates the accuracy or certainty in prediction
homogenisation effect of mining systems. Case studies for grade and the variability of the key quality parameter calorific value
control, stockpiling and blending strategy optimisation in diffuse (CV) of the shipped coal as a function of the operating mode.
deposits demonstrate the benefits the application of conditional The study is based on 25 realisations generated by conditional
simulation can generate (eg Benndorf and Dimitrakopoulos, 2010, simulation of the spatial distribution of the CV. A mining unit of
in this volume). 1500 t is considered, which reflects a train load.
This contribution investigates, based on conditional simulation, Figure 2 compares the digging mode of the two excavator
the ability to homogenise coal quality parameters in the process of types (chain and bucket wheel). Typically the bucket wheel
digging, conveying and stocking in a continuous mining excavator is positioned on a bench and digs the coal above the
environment. Objects of study are, in detail, the different excavator bench in a high cut mode in horizontal slices of about 5 m. This
types (bucketwheel excavator and chain excavator) as well as the enables a selective extraction of the typically horizontally
mode of operation of a strata type stockpile. A case study is deposited coal strata. The selectivity is limited by a minimum
presented in a lignite mine in Eastern Europe and conclusions mining thickness. In a deep cut mode operation, the chain
follow. excavator digs the face in vertical slices along the whole cutting
depths. This causes a blending effect of the horizontally
DESCRIPTION OF THE CONTINUOUS MINING deposited coal strata while digging, and limits its ability to
SYSTEM UNDER STUDY selectively mine this strata. Figure 3 shows photos of a bucket
wheel and chain excavator.
Figure 1 illustrates the continuous mining system considered in To analyse the variability of the coal quality parameters, block
this case study. The system contains two excavators, a chain and models were generated, which reflect the digging mode of both
a bucket wheel, which are positioned in different cuts at separate excavators.
belt conveyors. Both conveyors combine to join a central
stationary conveyor before the coal is loaded onto trains. There is
also the option to stack the coal in a coal stock and blending yard ANALYSIS OF THE ABILITY TO CONTROL COAL
with the goal being to homogenise the variability of coal quality QUALITY PARAMETERS WITHOUT UTILISING
parameters before shipping it. THE BLENDING YARD
1. GAusIMM, MIBRAG MBH, Germany. This section focuses on the described operation modes 1 to 3. It
Email: [email protected] is assumed that the mining equipment is successively digging the
J BENNDORF
Digging mode of different excavator types
Chain excavator
Bucket wheel excavator
FIG 2 - Digging modes of the two considered excavator types (bucket wheel and chain excavators).
FIG 3 - Photos of (left) a bucketwheel excavator and (right) a chain excavator (MIBRAG, 2009).
Prediction of the CV value of coal extracted by a bucket-wheel-excavator Prediction of the CV value of coal extracted by a chain-excavator
14,0 16,0
12,0 14,0
Calorific value (CV) in MJ/kg
10,0 12,0
8,0 10,0
6,0 8,0
4,0 6,0
1.000 6.000 11.000 16.000 21.000 26.000 31.000 36.000 1.000 6.000 11.000 16.000 21.000 26.000 31.000 36.000
Cumulated extracted tonnage Cumulated exctracted tonnage
FIG 4 - Predicted alorific value (CV) of coal delivered from the first FIG 5 - Predicted CV of coal delivered from the second cut utilising
cut utilising a bucket wheel excavator. a chain excavator.
blocks in the digging mode explained in the previous section at a operating mode 2, that is the chain excavator, in the second cut.
constant digging rate. Using the generated block models, the CV The figures show the prediction based on 25 realisations (thin
of the extracted coal is predicted. Figure 4 shows the predicted gray lines), the average of the realisations (thick line) as well as
CV for operating mode 1 that is the bucketwheel excavator in the the accuracy (certainty) in prediction by the means of the 90 per
first cut. Figure 5 shows respectively the predicted CV for cent confidence level (dashed lines).
INVESTIGATING THE VARIABILITY OF KEY COAL QUALITY PARAMETERS IN CONTINUOUS MINING OPERATIONS
The comparison of both figures illustrates the abilities and Vorhersage des Heizwertes
Prediction of the CV of(roh)
coalbeim
whengemeinsamen Betrieb
operating both beider Geräte
excavarors
limits of both types of equipment in selectively extracting and 14,00
homogenising the in-place variability of coal quality parameters. 13,00

When operating the bucket wheel excavator, a cyclic behaviour of
12,00
the CV can be detected. The reason is the combination of a
MJ/kg
strongly vertically varying CV and the vertical digging of slices
MJ/kg
11,00
in in
per digging location (set up). Due to the strong horizontal
value (CV)
10,00
CalorificHeizwert
correlation of the CV, a very similar cycle is repeated in the next 9,00
set up of the equipment. When operating a chain excavator, the
8,00
strong varying vertical CV values are homogenised over the whole
digging depth. The resulting flow of extracted coal is characterised 7,00
by a low variation of CV. 6,00

1500 6000 10500 15000 19500 24000 28500 33000 37500
Two criteria are used herein to evaluate the homogenisation Kumulativ
Cumulatedgeförderte Tonnage
extracted tonnage
effect in the different operating modes: the accuracy or certainty in

the prediction of a unit of 1500 t (train load), measured by the FIG 6 - Predicted CV of coal delivered from both excavators
standard deviation of the 25 realisations, and the difference operated at the same time.
between two successive extracted units as an expression of
variability. The difference in CV of successive delivered units is wheel excavator. On the other hand, this difference increases when
typically a criterion, which is part of the customer’s specification. compared to only operating the chain excavator.
Table 1 documents the results in terms of the criteria for the two
considered cases. The differences in the uncertainty of predicting To generalise, an operation of multiple excavators at the same
the CV between the two equipment types are due to the digging time decreases uncertainty in predicting the coal quality
mode, but also due to different areas of the mined deposit with parameters of shipped units. At the same time, extreme
different structural behaviour. However, the bucket wheel variability is smoothed. This homogenisation reduces differences
excavator shows significantly higher values of the uncertainty in between two successive extracted units. With the goal of
predicting the CV compared to the chain excavator. The reason is selectively extracting coal with a certain specification, the bucket
the occurrence of local blocks, which are due to the exploration wheel excavator should be operated separately.
level less certain in the prediction of CV. Due to the selective
extraction with the bucket wheel excavator, these blocks cannot be ANALYSIS OF THE ABILITY TO CONTROL COAL
blended with more certain blocks, as is possible with the chain QUALITY PARAMETERS UTILISING THE
excavator. Considering the second criteria, the difference between BLENDING YARD
two successive extracted units, the selectivity of the bucket wheel
excavator is clearly illustrated. With an average of approximately The key function of a coal stock and blending yard is, besides the
0.72 MJ/kg between the 1500 t units, the difference is about 80 per buffering, the homogenisation of coal quality parameters. This
cent higher compared to the chain excavator. The maximum transforms the characteristics of the incoming material flow in an
difference can be up to 7.95 MJ/kg. outgoing material flow whose characteristics are defined by
Figure 6 shows the prediction of the CV when operating both customer specifications and may be of contractual relevance
excavators at the same time (operating mode 3). Thereby with (Figure 7). The characteristics of the incoming material flow are
3000 m3/h for the bucket wheel excavator and 1500 m3/h for the a function of the geological conditions, the applied selectivity in
chain excavator, the two different digging rates are taken into extracting the deposit, and the operation mode in the pit.
account. The results in terms of the two criteria, uncertainty in The following considerations concern the homogenisation
prediction of a unit of 1500 t (train load) and the difference effect. The efficiency of blending and smoothing variability is
between two successive extracted units, are shown in Table 1. The significantly dependent on constructive parameters as well as the
blending of the two material flows from both excavators shows operation of the blending yard. Constructive factors are hereby the
two effects. The uncertainty in predicting the CV of the shipped type of the yard and its length and width. The following
coal decreases compared to only operating one of the two constructive parameters are assumed in the case study: the
excavators. On average, the standard deviation decreases by six blending yard is of type ‘strata’ (Figure 8). The coal is stacked
per cent compared to the chain, and 16 per cent compared to the layer by layer in six sub yards, which are each 120 m in length and
bucket wheel excavator. The second effect is a reduction in 50 m in width. Figure 9 illustrates the mode of operation of a
variability between two successive units compared to the bucket- ‘strata’ blending yard. The coal coming from the pit (incoming
TABLE 1
Summary of the criteria for measuring uncertainty in prediction of a unit of 1500 t (train load), and the difference between two
successive extracted units for different operating modes.
Information in this table are of unit MJ/kg Bucket wheel excavator Chain excavator Operating both
(1. cut) (2. cut) excavators
Certainty of predicting Average standard deviation 0.71 0.66 0.62
the CV value of 1.500 t
Min standard deviation 0.26 0.27 0.41
Max standard deviation 1.67 1.05 1.00
Average 90 per cent confidence level 1.87 1.92 1.76
Min 90 per cent confidence level 0.74 0.83 1.29
Max 90 per cent confidence level 4.41 2.99 2.90
Differences in CV Average differences between two successive 0.72 0.40 0.60
between successive units
extracted units
Min difference between two successive units 0.00 0.00 0.00
Max difference between two successive units 7.95 1.84 3.30
J BENNDORF
in-coming out-going
material flow Blending yard material flow
(from pit) (delivery)
FIG 7 - Process of blending in a coal stock and blending yard.
Reclaimer
Stacker
FIG 8 - Coal stock and blending yard of type strata (MIBRAG, 2009).
material flow) is stacked into layers, which are spread along the Description of the blending yard model and
total length of a subyard by a continuously up and down moving assumptions
stacker. The number and thickness of the several layers is variable
and can be influenced by the moving rate of the stacker dependent The model is based on the prediction of the CV of the extracted
on the total production rate of the mine. The yard is reclaimed in a coal while operating both excavators (operation mode 3) using
fan-shaped manner orthogonally to the alignment of the stacked 25 simulated realisations of the block model. Every single
layers by a scraper. In this way the coal quality of the outgoing realisation can be seen as a possible scenario of the coal quality
material flow is formed as an average over the total number of parameter CV of the produced material flow coming out of the
stacked layers. pit and being stacked into the blending yard. Figure 10 illustrates
the model of the blending yard. The idea is based on a study
The number of total layers in the yard can be varied by the shown in Benndorf (2009), and Pavloudavis and Agioutandis
mode of operation. The effectiveness of homogenisation shall be (2001). The model is composed of n columns and m rows. The
investigated subsequently as a function of the number of layers. number of layers in the yard is m. The single pockets are of size
Two criteria are used to evaluate the results: of 300 t and represent a fraction of the considered units of 1500 t.
1. the effect of homogenisation, which is defined herein as the Based on the assumptions regarding the dimensions of a
ratio between the dispersion variance of the CV of the subyard of 120 m by 50 m and an angle of repose of 35° of the
incoming material flow sin and the dispersion variance of the stacked material, the capacity of a subpile is about 50 000 t. With
CV of the out coming material flow sout (Equation 1) and; a pocket size of 300 t, the pile consists of 167 pockets or cells.
Depending on the number of layers m, the number of cells in a
2. the expected improvement of certainty in predicting the CV
layer is n. The assignment of CV to each cell reflects the incoming
of a unit of 1500 t, which is defined herein as the ratio
material flow and the up and down movement of the stacker. All
between the average standard deviation in predicting units of odd-numbered layers are stacked from left to right, and all
1500 t before stacking, and the average standard deviation in even-numbered layers are stacked from right to left. The
predicting units of 1500 t after reclaiming (Equation 2). reclaiming process occurs column by column. Because of the
constant cross section area of the pile, the necessary length to be
S in reclaimed for a unit of 1500 t is constant. Depending on the length
effect of homogenisation = (1)
S out of a cell in the pile, multiple columns can contribute to reclaim
one unit of 1500 t, or only a part of a column contributes to
reclaim one unit. The described model is implemented using VBA
σ 2 out for Excel and is used to investigate the homogenisation effect of
improvement of certainty in prediction = (2)
σ 2 in the coal stock and blending yard.
INVESTIGATING THE VARIABILITY OF KEY COAL QUALITY PARAMETERS IN CONTINUOUS MINING OPERATIONS
Width of the subyard
FIG 9 - Mode of operation of a blending yard of type strata.
Reclaiming
CVRmS1 CVRmS2 CVRmS... CVRmSn-1 CVRmSn

Stacking
CVRm-1S1 CVRm-1S2 CVRm-1S... CVRm-1Sn-1 CVRm-1Sn
Stacking
CVR...S1 CVR...S2 CVR...S... CVR...Sn-1 CVR...Sn
CVR2S1 CVR2S2 CVR2S... CVR2Sn-1 CVR2Sn
CVR1S1 CVR1S2 CVR1S... CVR1Sn-1 CVR1Sn
FIG 10 - Model of the coal stock and blending yard.
Analysis of the effect of homogenisation in realisation. In addition, the average of the realisations and a 90 per
dependence on the number of stacked layers cent confidence level is shown.
The following analysis is based on the results of the prediction of There is a clear difference between incoming and outgoing
CV, when operating two excavators (operating mode 3). The CV, which is the effect of the homogenisation of the blending
application of the 25 realisations of the CV of the incoming yard. Two effects can be detected:
material flow to the previously described model of the blending
1. the blending yard strongly transforms the varying CV of the
yard results in 25 realisations of the CV of the outgoing material
incoming material flow to a very homogeneous flow and
flow. Figure 11 shows the incoming material flow, while Figure 12
hardly disperses the CV of the outgoing material flow, and
shows the outgoing material flow for the case of 15 layers in the
pile. The figures show the predicted CV per 1500 t unit for each 2. a strong improvement of the certainty of predicting the CV.
J BENNDORF
Predicted coal quality (CV) of the in-coming material flow similar homogenisation effect. This has implications on the
14,00 operation mode of the blending yard. The number of stacking
13,00
passes can be reduced from 61 to 35, which cause a reduction of
operational costs and increases the safety level in the operating
12,00
area. For the improvement of predicting CV, a similar observation
11,00 can be done. The standard deviation of units of the outgoing
material flow is about 50 per cent of the incoming material flow.

10,00
This effect can also be achieved with 35 layers. The average
9,00 absolute value of the standard deviation of predicted CV of the
8,00
outgoing 1500 t units is ± 0.30 to ± 0.35 MJ/Kg. This represents
the standard deviation of predicting a block of 50 000 t in the
7,00
deposit that means a block of the size of a subpile in the blending
6,00 yard. In general, the standard deviation of predicted CV of the
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34
Extrated units of 1.500t
1500 t units in the outgoing material flow converges to the
standard deviation of the predicted CV of the whole subpile. This
FIG 11 - Predicted CV of the incoming 1500 t coal units.
observation is essential in designing subpiles and defining its size
as a function of the exploration level of the deposit and the
customer’s specifications.
Predicted coal quality (CV) of the in-coming material flow (after reclaiming)
14,00
13,00
SUMMARY AND CONCLUSIONS
12,00 The properties of the conditional simulation in geostatistics allow
11,00
for investigating the in situ variability of coal quality parameters
and its behaviour along the extracting, conveying and blending

10,00
process. The case study illustrated an application in a continuous
9,00 mining environment with different types of excavators and the
8,00
utilisation of a coal stock and blending yard. The results deliver
valuable insight into the effect of homogenisation and an
7,00 improvement in predicting CV values of shipped coal units as a
6,00 function of the operational mode. Different excavators cause a
1 3 5 7 9 11 13 15 17 19 21 23
Reclaimed units of 1500t P(train loads)
25 27 29 31 33 different behaviour of CV of the extracted coal. Bucketwheel
excavators cause substantial differences between extracted units,
which show a cyclic behaviour. Chain excavators tend to smooth
FIG 12 - Predicted CV of the outgoing 1500 t coal units.
out CV and show a homogenisation effect. When operating both
excavators simultaneously, an improvement in predicting CV can
The homogenisation effect of the blending yard is due to the be achieved. Furthermore, the advantage of utilising a strata-type
elimination of successive auto-correlated units. It is highly stockpile in the conveying process was demonstrated by a
probable that a unit of high CV is followed by another unit of a substantial improvement in predicting and homogenisation CV.
high CV in case of the incoming material flow. This correlation With a goal of meeting production goals with a high probability,
is disconnected in the outgoing material flow, which is caused by as well as delivering a homogeneous product, an operation of
the fan-shaped reclamation of the layered stacked coal. excavators without utilising the stockpile should be avoided. The
To investigate the homogenisation effect, the above study is results are a basis for optimising the operation and adjusting the
performed on a varying number of layers in the pile. The number operational mode day-to-day, depending upon the customer’s
takes values in an interval from one to 61, which defines the requirements and the current geological setting.
technical limits of the blending yard under consideration. Results
in terms of the homogenisation effect and improvement of REFERENCES
certainty in predicting CV of shipped 1500 t units are presented
in Figure 13. Benndorf, J, 2009. Evaluation of lignite deposits using conditional
simulation in geostatistics, Schriftenreihe des Institutes für
Geotechnik und Markscheidewesen der TU Clausthal, Heft 18/2009,
Effect of homogenisation and improvement in certainty in prediction 221 p (in German).
20,0 Benndorf, J and Dimitrakopoulos, R, 2010. Stochastic long-term
1,00 18,0 production scheduling of iron ore deposits – Integrating joint
16,0 multielement geological uncertainty, in Advances in Orebody
Improvement in prediction
Modelling and Strategic Mine Planning I (ed: R Dimitrakopoulos),

effect of homogenization
0,80
14,0
12,0
0,60 Melbourne).
10,0
Costa, J F, Zingano, A C and Koppe, J C, 2000. Simulation – An
8,0
0,40 approach to risk analysis in coal mining, Exploration and Mining
6,0
Geology, 9(1)43-49.
4,0
0,20
Improvement in prediction Effect of homogenization
Goovaerts, P, 1997. Geostatistics for Natural Resources Evaluation
2,0
(Oxford University Press: New York).
0,00 0,0
1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 35 41 45 51 61
Hohn, M E and McDowell, R R, 2001. Stochastic simulation of coal bed
Number of stacked layers thickness and economic decision-making, Geologic Modelling and
Simulation: Sedimentary Systems, (eds: D F Merriam and J C Davis),
pp 271-283 (Kluwer Academic/Plenum Publishers: New York).
FIG 13 - Homogenisation effect and improvement of certainty in
MIBRAG, 2009. MIBRAG web site [online]. Available from: <www.
predicting CV as a function of layers in the blending yard.
mibrag.de> [Accessed: September 2009].
Pavloudakis, F and Agioutantis, Z, 2001. Development of a software tool
In the case study, the dispersion between the delivered coal units for the prediction of the coal blending efficiency in longitudinal
can be reduced to a 17th of the incoming material flow by utilising stockpiles, in IMCET 2001, 17th International Mining Congress and
all 61 layers. However, 35 layers appear to be sufficient for a Exhibition of Turkey, Ankara, Turkey, 19 - 22 June.
HOME
Cut-Off Grade Estimation – Old Principles Revisited – Application

to Optimisation of Net Present Value and Internal Rate of Return
J-M Rendu1
ABSTRACT grade optimisation is unavoidable. Decisions must be made using

The modern principles of cut-off grade estimation were first published by
complex objectives or combinations of potentially conflicting
Ken Lane over 40 years ago. These principles form the basis of most objectives. Fundamental principles are first revisited, which give
computerised systems routinely used in mine design and production a framework within which quantifiable and occasionally not
scheduling. In this paper, the author summarises these principles and quantifiable variables can be taken into account. How these
proposes simple equations to estimate cut-off grades in a variety of principles can be used to make increasingly complex decisions is
situations. All consequences of choosing a cut-off grade are discussed illustrated with financial objectives which include optimisation
including financial, technical, economic, legal, environmental, social and of cash flow, of Net Present Value and of Internal Rate of Return.
political ones. A utility function is proposed as a means to take into
account various potentially conflicting objectives. An example is given
which illustrates how cut-off grades can be calculated to maximise Net UTILITY FUNCTION
Present Value (NPV) and Internal Rate of Return (IRR). The complex
relationship between naturally fixed geologic parameters and parameters Choosing a cut-off grade is equivalent to choosing the value xc of
on which humans have some control is discussed. a geologically defined parameter (or set of parameters) x which
will be used to decide whether one tonne of material that becomes
INTRODUCTION available for processing in a particular period of time should be
sent to one process or another. In simple cases, a single parameter
A cut-off grade is generally understood to be the value of a may be sufficient to define the destination of material being
geologically defined parameter used to separate material which mined, such as a copper grade or a gold grade. In other cases, a set
should be wasted from that to be sent to the processing plant. In of parameters may have to be considered such as copper and
the simplest situations, the cut-off grade is a minimum metal molybdenum grades, sulfide content, clay content, mineralogy and
content below which material is sent to the waste dump and percentage of deleterious elements and the cut-off ‘grade’ must be
above which it is sent to the mill. Cut-off grades are also used to defined as a function of these parameters. The term ‘utility’ is used
decide the routing of mined material when two or more processes here as it is in decision theory (Raiffa and Schlaifer, 1961) to
are available, such as heap leaching and milling. Cut-off grades represent the satisfaction gained from following a given course of
are used to decide whether material should be stockpiled for action. This satisfaction is a function of preferences and values
future processing or processed immediately. Many publications specific to the decision maker. The utility of a given cut-off grade
can be found in the technical literature concerning estimation strategy is a measure of the extent to which this strategy reaches
and optimisation of cut-off grades, the most comprehensive the mining company’s objectives. The utility of sending one tonne
being the reference book by K F Lane, The Economic Definition of material with parameter value (grade) x to destination 1
of Ore: Cut-off Grades in Theory and Practice (Lane, 1988). An (Process 1) is U1(x). The utility of sending the same material to
introduction to the subject can be found in Rendu (2008). destination 2 (Process 2) is U2(x). The cut-off grade xc is the value
This paper attempts to clarify the many consequences of of x for which:
changing cut-off grades. Cut-off grades are calculated by U1(xc) = U2(xc)
comparing costs and benefits. A cut-off grade strategy defines
the properties of the material being sent to the processing plant, If U1(x) exceeds U2(x) for x greater than xc then all material
including tonnage, average grade and metallurgical properties. for which x is greater than xc should be sent to Process 1. The
Expected cash flows are function of this strategy. Most cut-off choice of a cut-off grade is governed primarily by financial
grade optimisation algorithms are based on the assumption that objectives. However the consequences of choosing a given
the owner’s objective is to maximise the NPV of future cash cut-off grade are complex and not all of a financial nature. A
flows, NPVi calculated using a specified discount rate i. variety of possibly conflicting objectives must be taken into
To reach this objective one must consider two classes of account. To illustrate the role of these different objectives in the
variables: decision making process, the utility function U(x) is written as
the sum of three parts:
• Class A – relating to the properties of the naturally occurring
mineralisation, and U(x) = Udir(x) + Uopp(x) + Uoth(x)
• Class B – relating to human-driven economic constructs and Udir(x) represents the direct, undiscounted profit or loss which
technology. will result from processing one tonne of material of grade x.
Class A variables are defined by nature and include the shape Uopp(x) represents the opportunity cost or benefit of changing the
of the deposit and its geologic and metallurgical properties. Class processing schedule by adding one tonne of grade x to the
B variables are at least partly the result of engineering, economic material flow. When there are capacity constraints, adding a new
and other factors that control mining sequence, costs, revenues tonne of material to a predetermined processing schedule cannot
from sales and ultimately cash flows. The complex relationship be done without changing this schedule. Such a change may have
between Class A and Class B variables must be taken into positive or negative consequences which are incorporated in the
account to optimise cut-off grades. The complexity of cut-off opportunity cost. Opportunity costs occur only when there are
constraints, which limit how many tonnes can be mined or
processed at a given time. Other factors that must be taken into
1. FAusIMM, Mining Consultant, 5451 South Geneva Street, Englewood account in the calculation of cut-off grades, but may not be
CO 80111, USA. Email: [email protected] quantifiable, are represented by Uoth(x).
J-M RENDU
DIRECT PROFIT AND LOSS function of both mining and milling costs and is known as the
mine cut-off grades:
The marginal cut-off grade is calculated taking only direct
operating costs and corresponding revenues from sales into xc = [Mo + Po + Oo]/[r × (V - R)]
account, if there are any. Capacity constraints are ignored. Cash
flows are not discounted. Opportunity costs are not taken into OPPORTUNITY COSTS AND BENEFITS
consideration, and neither are any other consequences which
changing the cut-off grade may have on the economics of the Opportunity costs Uopp(x) may result from mining and processing
project. Only costs that can be directly attributed to the mining one tonne of material that was not previously scheduled to be
and/or processing of one tonne of material are taken into mined or processed. No opportunity cost is incurred if mine, mill
account. This may include some costs classified as overhead, to or refining facilities are not capacity constrained – only direct
the extent that such costs will change with the life or size of the costs and benefits need be taken into account. If there are
operation. The marginal cut-off grade is calculated by reducing capacity constraints, adding a new tonne to the production
the utility function to its first term: schedule will result in displacing material already scheduled to
be processed and postponing treatment of this material. The cost
U(x) = Udir(x) of displacing this material is the opportunity cost. The
The utility of mining and processing one tonne of material of opportunity cost must be added to the direct cost of the process
grade x is: that is capacity-constrained. If the mine is capacity-constrained,
the opportunity cost applies to both ore and waste and has no
Uore(x) = x × r × (V - R) – (Mo + Po + Oo) impact on cut-off grade, unless waste material can be left in the
where: ground, as may be the case at the bottom of an open pit or along
the walls of an underground stope. If the mill is capacity-
x represents the average grade constrained the opportunity cost is added to all tonnes processed.
r is the proportion of valuable product recovered from the Consider a project whose production schedule has been
mined material designed to fill up the processing plant to capacity. The Net
Present Value of future cash flows is NPVi. This Net Present
V represents the value of one unit of valuable product
Value was calculated on the basis of currently planned
R stands for refining costs that are related to each unit of production using a discount rate i. If we add one new tonne of
valuable material produced material to the capacity-constrained processing plant we are
Mo represents the mining cost per one tonne of ore postponing treatment of the originally scheduled material by the
time t needed to process the additional tonne. This time is equal
Po the processing cost per one tonne of ore to the inverse of the plant capacity. Adding one new tonne today
Oo the overhead cost per one tonne of ore will decrease the Net Present Value of previously scheduled
future cash flows by t × i × NPVi. The opportunity cost of adding
The utility of mining and wasting the same material is: one tonne of material to a capacity-constrained operation is:
Uwaste(x) = - (Mw + Pw + Ow)
Uopp(x) = - t × i × NPVI
where:
Because the Net Present Value of future cash flows decreases
Mw represents the mining cost per one tonne of waste over time, the opportunity cost also decreases and the optimal
Pw represents the processing cost per one tonne of waste that cut-off grade increases. At the end of the mine life the
may be needed to avoid potential water contamination and opportunity cost is zero and the applicable cut-off grade is the
acid generation minimum cut-off grade calculated using only direct costs. The
above formula is useful to verify that cut-off grades and Net
Ow stands for the overhead cost per one tonne of waste Present Value have been optimised. However cut-off grades
To decide whether one tonne of material should be wasted or calculated from cash flows which have not been optimised are
sent to the processing plant, one must compare the utility of both also not optimal and an iterative approach must be used to
actions. The minimum cut-off grade is the value xc of x for estimate the cut-off grade. For example one can first calculate a
which: cash flow using a fixed cut-off grade such as that calculated
without opportunity cost. The Net Present Value of this cash flow
Uore(xc) = Uwaste(xc)
can be calculated and the theoretically optimal cut-off grade can
xc = [(Mo - Mw) + (Po - Pw) + (Oo - Ow)]/[r × (V - R)] be re-estimated using the corresponding opportunity cost. But
new cut-off grades imply new mine plans, new cash flows, and
In this formula, the numerator represents the difference therefore new opportunity costs. These updated opportunity costs
between direct costs incurred when treating the material as ore must be used to re-estimate the cut-off grades. If the theoretically
and those incurred when treating it as waste. The denominator optimal cut-off grades are significantly different from those
represents the revenues resulting from the sale of one unit of which were used in the Net Present Value calculation, significant
contained valuable product. If the costs of mining and shipping successive modifications to the mine plan may be needed before
material to the waste dump or to the processing plant are the an optimum solution is found.
same (Mo = Mw) and there are no significant additional costs in
processing waste (Pw = 0 and Ow = 0), this cut-off grade is only a Algorithms can be found in the technical literature and
computer programs have been developed to facilitate cut-off grade
function of mill costs and recoveries and is independent of
optimisation. Because of the complex relationship between the
mining costs. It is known as the mill cut-off grades: geological properties of the deposit and human-driven parameters
xc = [Po + Oo]/[r × (V - R)] which define production schedule and cash flow, no simple
solution exists while improved algorithms and easier to use
If the material considered need not be mined because it is at computer programs are continuously developed. Examples of
the bottom of the pit or along the walls of a stope, the utility of proposed solutions can be found in publications by Carter, Lee,
treating it as waste is Uwaste(x) = 0. The material will be mined and Baarsma (2004), Chanda (2004), Dagdelen and Kawahata
and processed only if Uore(x) is positive. This cut-off grade is a (2008), Dimitrakopoulos, Martinez and Ramazan (2004), Hall
CUT-OFF GRADE ESTIMATION – OLD PRINCIPLES REVISITED
(2003), Hoerger, Bachmann, Criss, and Shortridge (1999), disadvantages of each project with the subjective cost/benefit
Hoerger, Hoffman, and Seymour (1999), King (2001), Lerchs and to all stakeholders, according to company-specific values and
Grossmann (1965), Lane (1988), Whittle (1999) and others. priorities.
OTHER COSTS AND BENEFITS DIFFERENCE BETWEEN MAXIMISING NPV

Cut-off grades play a critical role in defining all major AND MAXIMISING IRR
characteristics of a mine, including the size of operation, capital The choice between business objectives is a rational decision
and operating costs, life of the mine and cash flow. In addition to based on the environment within which the company operates,
the economically quantifiable financial impact that cut-off grade including the availability of mutually exclusive and/or independent
changes may have, there are other costs and benefits which are investment opportunities and the extent to which capital resources
often not easily quantifiable but must be taken into account. are constrained. Optimisation of Net Present Value is the objective
Consideration must be given not only to changes in cash flow and most commonly used to calculate cut-off grades. In some
Net Present Value as measured by Udir(x) and Uopp(x), but also to situations management’s primary objective might be maximisation
all other impacts Uoth(x), including those of an environmental, of the Internal Rate of Return (IRR).
socio-economic, ethical and political nature. All stakeholders must Consider the case where a feasibility study has been completed
be taken into account to evaluate different cut-off grade strategies. on a new project. The capital cost of bringing the project to
These stakeholders include: production is I. Production schedules have been developed and
• shareholders; corresponding cash flows have been estimated. Using a discount
rate i, the Net Present Value of expected future cash flows is
• banks and other financial institutions; NPVi. This Net Present Value is a function of the point in time
• regulators; when it is calculated. By definition the Internal Rate of Return
IRR is the value ir of i for which the Net Present Value of the
• employees and their families; project, calculated before the investment is made and therefore
• customers who buy the product sold by the company; including the capital investment I, is zero:
• suppliers of goods and services; NPVir = 0
• local communities; If there was no capacity constraint, maximum rate of return and
• local, regional, federal and country governments; maximum Net Present Value would both be obtained by
maximising production over the shortest possible time period. In
• non-governmental agencies (NGOs); and
practice, except for occasional periods during the life of a mine, or
• future generations which will live with the long-term when mine and mill have been over-designed, all mining
environmental impact of the mining operation. operations are capacity constrained. Such constraints are imposed
Management decides how to balance financial goals with other by the need to balance capital cost, operating cost and production
stakeholder requirements and expectations. The success of a schedule. Without capacity constraint, the optimum cut-off grade
project and the company’s long-term public image and licence to xc1 between material wasted and that sent to the mill is:
operate depend on skilful management of potentially conflicting xc1 = [(Mo - Mw) + (Po - Pw) + (Oo - Ow)]/[r × (V - R)]
objectives. Practical guidelines must be given to those in charge
of mine planning to ensure that the projects are designed to reach If the mill is capacity-constrained, adding one tonne to the
the company’s objectives. production schedule results in an opportunity cost Uopp(x):
Higher cut-off grades may increase short term profitability,
Uopp(x) = - t × i × NPVi
enhance return to shareholders, shorten the payback period and
reduce political risk. However, for potentially long-life projects, A higher cut-off grade must be used, xc2, calculated as follows:
significantly shortened mine life can reduce time-dependent
opportunities such as those offered by price cycles. Lower cut-off xc2 = [(Mo - Mw) + (Po - Pw) + (Oo - Ow) + t × i × NPVi]/
grades may increase project life, with potentially more [r × (V - R)]
sustainable economic impact. Lower cut-off grades imply fuller
consumption of mineral resources, which may be politically This is the optimum cut-off grade when the primary objective is
advantageous or legally required. However, lower cut-off grades to maximise the Net Present Value using the discount rate i. In this
may also result in higher sensitivity to price and cost variations. formula, the Net Present Value NPVi must be calculated from
Stakeholders may have to choose between higher returns over expected future cash flows, beginning with the period for which
short time periods or low returns over longer time periods. In the cut-off grade is calculated. The cut-off grade that applies to
some circumstances, using high but decreasing cut-off grades year one must be calculated using the Net Present Value of cash
early in the mine life and stockpiling low-grade material for later flows beginning with year one. This Net Present Value is a
processing can help balance economic uncertainty, financial function of the production schedule but is independent of any past
returns, resource consumption and mine life. investment. Past investments enter into the calculation indirectly
A method that can be used to optimise cut-off grades while and they enter only because of capacity constraints that were
taking into account costs and benefits which cannot be imposed by investment decisions made at an earlier date. Because
quantified, consists of evaluating the project under a variety of the Net Present Value of future cash flows usually decreases when
constraints imposed on mine capacity, mill capacity, volume of the remaining mine life decreases, the cut-off grade also decreases
sales, capital or operating costs, life of mine, etc. A different over time.
cut-off grade profile is developed under each set of constraints. If maximisation of Internal Rate of Return is the primary
These cut-off grades are chosen to optimise Net Present Value or objective, the desired discount rate is ir as defined above. The
Internal Rate of Return. The relationship between discount rate formula to be used to calculate the optimal cut-off grade can be
and Net Present Value, the project Internal Rate of Return, its re-written as follows:
sensitivity to cost and price changes and risk analysis, are some
of the measures used to compare projects. The final development xc3 = [(Mo - Mw) + (Po - Pw) + (Oo- Ow) + t × ir × NPVir]/
decision is made by comparing the quantified advantages and [r × (V - R)]
J-M RENDU
In this formula, ir is the calculated Internal Rate of Return and expected yearly cash flow and Net Present Value of future cash
is a function of the capital investment I. At the beginning of the flows discounted at ten per cent and 17.2 per cent respectively.
first year of production the Net Present Value NPVir is equal to Two modifications to the mine plan are being considered – the
the total initial investment I, updated to year one, taking into objectives being to optimise the Net Present Value at a ten per
account the discount rate i in the calculation of the present value cent discount rate and to maximise the Internal Rate of Return,
of past investments. The same formula applies to the calculation respectively. The only constraint imposed on these modifications
of cut-off grade, whether the objective is to maximise NPV or is that the mill capacity must remain at 36 million tonnes. A first
IRR, except that in the first case the discount rate is fixed while step consists in calculating the theoretically optimal cut-off grade
in the second case it changes with the size of the initial for Year 1. To do this, one calculates the Net Present Value of
investment. future cash flows as of the beginning of Year 1:
When considering a new project, different development options • NPV10% = $2 130 000 000
must be considered. Different mine and mill capacities are
considered, resulting in different production schedules, cash flows, • NPV17.2% = $1 500 000 000
and capital requirements. Each one of these options corresponds to If there was no capacity constraint the minimum cut-off grade
a different Internal Rate of Return and optimal cut-off grade would be:
profile. For each option, the Internal Rate of Return is a function
of planned production, which in turn is determined by the cut-off xc1 = [(Mo - Mw) + (Po - Pw) + (Oo - Ow)]/[r × (V - R)]
grade. Given an estimated cash flow, one can calculate a
theoretically optimal cut-off grade profile. If this theoretical = [(1.45 - 1.45) + (4.00 - 0.05) + (0.50 - 0.05)]/[0.86 ×
optimum is different from the cut-off grade actually used to (1.20 - 0.30) × 2205]
develop the production schedule, changes to the cut-off grade
= 4.400/1707 = 0.26% Cu
should be considered.
In this formula, 2205 is the number of pounds per tonne. If the
Example objective is to maximise the Net Present Value NPV10% the Year 1
The following example illustrates how cut-off grades can be opportunity cost to be added to the processing cost is calculated
estimated depending on whether the objective is maximisation of as follows:
Net Present Value or of Internal Rate of Return. A feasibility Uopp(x) = - t × i × NPV10%
study has been completed on a proposed new copper mining
operation. The capital investment needed to bring the mine to = - (1/36 000 000) × 10% × 2 130 000 000
production has been estimated at $1.50 billion. The planned
mining capacity is 72 million tonnes per year (including both ore = -$5.917 per tonne mined
and waste), and the mill capacity is 36 million ore tonnes per
The theoretically optimal Year 1 cut-off grade is xc2:
year. The following operating parameters have been estimated:
r = 86% (including mill and smelter recoveries) xc2 = (4.400 + 5.917)/1707 = 0.60% Cu
V = $1.20 per pound of copper sold If the objective is to maximise the Internal Rate of Return the
Year 1 opportunity cost to be added to the processing cost is
R = $0.30 per pound of copper for freight, smelting and refining calculated as follows:
Mo = $1.45 per tonne of ore mined
Uopp(x) = - t × ir × I
Po = $4.00 per tonne of ore processed
= - (1/36 000 000) × 17.2% × 1 500 000 000
Oo = $0.50 per tonne of ore processed
= -$7.167 per tonne mined
Mw = $1.45 per tonne of waste
Pw = $0.05 per tonne of waste The theoretically optimal Year 1 cut-off grade is xc3:
Ow = $0.05 per tonne of waste xc3 = (4.400 + 7.167)/1707 = 0.68% Cu.

The feasibility study was based on a 14-year mine life and a The cut-off grades xc2 and xc3 are optimal only during the first
mill head grade of 0.91 per cent Cu. Using a ten per cent year of production, provided that Net Present Value and Internal
discount rate, the project Net Present Value is $630 000 000 and Rate of Return were calculated from mine plans which used
the Internal Rate of Return is 17.2 per cent. Figure 1 shows the these cut off grades. Figure 1 shows how the Net Present Value
of future cash flows decreases over time. Figure 2 shows how the
CashFlow NPV10% NPV17.2% cut-off grade, calculated using these Net Present Values,
2.5E+09
decreases over time.
This example illustrates why optimisation of cut-off grades is a
NPVofFutureCashFlow
2.0E+09 difficult, lengthy, iterative process. Cut-off grades calculated using

the above formulae are optimal only if the cash flows used to
1.5E+09
calculate these cut-off grades were developed using optimal
cut-off grades. In a typical situation initial assumptions will be
made concerning the mine and plant capacities which are
1.0E+09
technically feasible, consistent with geologic constraints and
expected to give a reasonable rate of return. Production schedules
5.0E+08
will be developed using cut-off grades not lower than the marginal
cut-off grade. Capital costs and operating costs will be estimated
0.0E+00 accordingly, from which cash flows will be developed. It is not
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 likely that the cut-off grades used in this initial production
Year schedule and cash flow analysis will be optimal. To reach
the optimum, one must consider changing the production
schedule, taking into account the calculated theoretically optimum
FIG 1 - Yearly cash flow and Net Present Value of future cash flows. cut-off grades. These changes are subject to geologic, technical
CUT-OFF GRADE ESTIMATION – OLD PRINCIPLES REVISITED
Nodiscount 10.0%discount 17.2%discount REFERENCES

0.8% Carter, P G, Lee, D H and Baarsma, H, 2004. Optimisation methods for
the selection of an underground mining method, in Proceedings
0.7%
Orebody Modelling and Strategic Mine Planning (ed: R
CutoffGrade(Cu%)
0.6% Dimitrakopoulos), pp 7-12 (The Australasian Institute of Mining and

0.5% Metallurgy: Melbourne).
Chanda, E K, 2007. Network linear programming optimisation of an
0.4% integrated mining and metallurgical complex, in Orebody
0.3% Modelling and Strategic Mine Planning, second edition (ed: R
Dimitrakopoulos), pp 149-155 (The Australasian Institute of Mining
0.2%
0.1% Dagdelen, K and Kawahata, K, 2008. Value creation through strategic
0.0% mine planning and cutoff-grade optimisation, Mining Engineering,
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
60:39-45.
Dimitrakopoulos, R, Martinez, L and Ramazan, S, 2007. Optimising open
Year pit design with simulated orebodies and Whittle Four-X: A
maximum upside/minimum downside approach, in Orebody
Modelling and Strategic Mine Planning, second edition (ed: R
FIG 2 - Yearly cut-off grades at various discount rates. Dimitrakopoulos), pp 201-206 (The Australasian Institute of Mining
and financial constraints. In the example above, supplying Hall, B E, 2003. How mining companies improve share price by
36 000 000 t to the processing plant may not be achievable if the destroying shareholder value, in Proceedings CIM Mining
Conference and Exhibition, paper 1194, (The Canadian Institute of
cut-off grade is raised to 0.68 per cent Cu. Significant changes in
Mining and Metallurgy: Montreal).
cut-off grade may require commensurate changes in mine
Hoerger, S, Bachmann, J, Criss, K and Shortridge, E, 1999. Long-term
production schedules. Mining constraints may be encountered
mine and process scheduling at Newmont’s Nevada operations, in
before the necessary tonnage of mill feed can be mined. These Proceedings 28th APCOM Symposium (ed: K Dagdelen), pp 739-747
constraints must be taken into account in addition to the (The Society for Mining, Metallurgy and Exploration Inc: Littleton).
processing constraints. When both mine and mill are capacity Hoerger, S, Hoffman, L and Seymour, F, 1999. Mine planning at
constrained, cut-off grades are determined by the geologic Newmont’s Nevada operations, Mining Engineering, 51:26-30.
properties of the deposit, including tonnage and grade of material King, B, 2001. Optimal mine scheduling, mineral resource and ore
which can be mined. Financial considerations become secondary. reserve estimation, in Mineral Resource and Ore Reserve Estimation
– The AusIMM Guide to Good Practice (ed: A C Edwards),
CONCLUSIONS pp 451-458 (The Australasian Institute of Mining and Metallurgy:
Melbourne).
The formulae presented in this paper are simple and easy to use. Lane, K.F, 1988. The Economic Definition of Ore: Cut Off Grades in
They are particularly useful to check whether the cut-off grades Theory and Practice, 149 p (Mining Journal Books Ltd: London).
planned for development of a new mine, or actually used in an Lerchs, H and Grossmann, I F, 1965. Optimum design of open-pit mines,
operating mine are optimal or nearly so. They are also useful to Transactions, LXVIII:17-24 (The Canadian Institute of Mining and
check whether production schedules developed using available Metallurgy: Montreal).
mine planning and scheduling software are reasonable. However, Raiffa, H and Schlaifer, R, 1961. Applied Statistical Decision Theory,
geologic environment and changes in operating and economic 356 p (Harvard University: Boston).
conditions can impose constraints which require a cut-off grade Rendu, J-M, 2008. An Introduction to Cutoff Grade Estimation, 112 p
significantly different from the theoretical optimum at any time. (The Society for Mining, Metallurgy and Exploration Inc: Littleton)
This difference should be analysed to make sure that it can be Whittle, J, 1999. A decade of open pit mine planning and optimisation –
explained and justified. A theoretical optimum is only an The craft of turning algorithms into packages, in Proceedings 28th
optimum if all constraints are taken into account. It is because of APCOM Symposium (ed: K Dagdelen), pp 15-23 (The Society for
Mining, Metallurgy and Exploration Inc: Littleton).
the complex relationship between fixed parameters, determined
by geology and parameters on which mine operators have some
control, that cut-off grade optimisation remains an elusive
problem which will keep many practitioners and researchers
challenged for years to come.
HOME
Using Grade Uncertainty to Quantify Risk in the Ultimate Pit

Design for the Sadiola Deep Sulfide Prefeasibility Project, Mali,
West Africa
S Robins1
ABSTRACT the LOM extended to 2024. SEMOS SA therefore initiated the

deep sulfide prefeasibility project to optimise the exploitation of
In order to quantify the uncertainty in the grade estimate for the Sadiola
deep sulfide prefeasibility project, a conditional simulation model was the hard sulfide component of its mineral resource. The
generated using the direct block simulation methodology. Compared to prefeasibility project focused on the pit expansion potential of
conventional sequential Gaussian simulation, the direct block simulation the deposit, as well as the options available to upgrade the
algorithm produced a reliable model in significantly less time, lending its current gold treatment plant to process the hard sulfide material.
application to a production environment. Through application of a mining As of June 2006, cost estimates indicated that a capital outlay of
transfer function, risk pits were generated for comparison with the deep approximately USD $145 million would be required to upgrade
sulfide prefeasibility pit. The results of this study revealed that the the treatment plant to efficiently process hard sulfide material.
prefeasibility pit is optimal at the applied gold price and cost parameters,
and that the risk of not achieving the project grade profile is low. Should
Given the large capital outlay required to commence with the
the gold price increase, or the operating costs of the project decrease deep sulfide Project, it was necessary to quantify the technical
significantly, the deep sulfide reserve tonnage would realise significant risk of exploiting the Sadiola hard sulfide resource.
upside potential. Probability and uncertainty analysis revealed that the Quantification of the grade uncertainty was vital to providing
greatest risk to the project is the confidence in the footwall grade focus areas that required additional drilling before a reliable
estimate. At a drill spacing of 50 m × 50 m and a sample interval of 1 m, feasibility study grade model could be obtained. The Sadiola
the probability of the footwall grade exceeding the economic cut-off of mineral resource model used for the prefeasibility study was
2.0 g/t is low, while the uncertainty in the grade estimate is high. comprised of a grade estimate, into 30 m × 30 m × 10 m blocks,
Although significantly lower in grade than the main zone, which is the
primary economic driver of the project, the footwall mineralisation is
using ordinary kriging. This was followed by a change of support
important in terms of reducing stripping ratio and delivering ore tonnes to calculation, using uniform conditioning, to an SMU support size
optimise the treatment schedule. This zone is therefore a focus area for of 10 m × 10 m × 5 m. The sample information on which this
further drilling. model was built was comprised predominantly of NQ (53 mm
diameter) half diamond core, with 4.5 inch RC precollars
INTRODUCTION through the soft material. The final model comprised
13 independent estimation domains to honour the stationarity
Sadiola Hill Gold Mine is located at latitude 13°56’N, longitude requirements for kriging.
11°40’W, and altitude 125 m above mean sea level, which places This model effectively generated a single grade outcome for
it approximately 500 km north west of Bamako, the capital of planning purposes. Questions remained as to how accurate this
Mali in West Africa (Figure 1). The mine is operated by the grade estimate was, and if there was a practical method for
Societe d’Exploration des Mines d’Or de Sadiola SA (SEMOS determining the uncertainty in the grade estimate. Conditional
SA), which comprises a joint venture partnership between simulation was utilised to answer these questions. Conditional
AngloGold Ashanti (38 per cent), IAMGOLD (38 per cent), the simulation, a Monte Carlo-type simulation approach, generates
Malian Government (18 per cent) and the IFC (six per cent).
multiple and equally probable realisations that provide a model
Construction of the Sadiola plant began in March 1995 and was
of the spatial uncertainty of the grade in the in situ orebody
completed in June 1997. Gold production began in mid 1997
(Dimitrakopoulos, Farelly, and Godoy, 2002). The resultant
from the saprolite oxide orebody and more recently from
realisations are not only conditioned to the available sample data,
saprolite sulfide ore as mining has progressed through the
weathered profile towards the hard/soft boundary. The Sadiola and therefore all reasonably match the same sample statistics, but
orebody extends substantially below the weathered horizons into also reasonably duplicate the histogram and semi-variogram
the underlying hard sulfide lithologies. However, the current Life model of the sample data (Goovaerts, 1997). Spatial features are
of Mine plan (LOM) does not exploit this material because the deemed ‘certain’ if they are present in most of the simulated
plant is unable to treat more than ten per cent hard material. In maps and ‘uncertain’ if seen on a few simulated maps.
2005, Sadiola treated approximately 5.3 Mt of ore at 2.83 g/t, According to Goovaerts (1997), generating alternative
accounting for a gold production of 483 koz (Van der realisations of the spatial distribution of an attribute is rarely the
Westhuizen, 2005). The Sadiola plant was designed to treat this goal. Rather these alternative realisations serve as the input to
material, comprised almost entirely of saprolite and laterite. At a other transfer functions, which in the open pit mining
cut-off grade of 1.42 g/t, the remaining Sadiola Main Pit Reserve environment would comprise a mining process such as a pit
(at the end of 2005) was 21 Mt at 3.35 g/t, accounting for optimisation algorithm. In assessing the uncertainty in the grade
2.3 Moz. This reserve would accommodate mining to the end of estimate for the Sadiola mineral resource model, the goal was to
2009. generate a number of different grade realisations and run each of
The hard sulfide portion of the Sadiola main pit mineral these realisations through a pit optimisation algorithm in order to
resource at the end of 2005 was 48.5 Mt at 2.25 g/t, accounting determine where the prefeasibility design pit ranked. The
for 3.5 Moz (at 0.70 g/t cut-off grade), providing large upside challenge lay in selecting a simulation method that was practical
potential for the mine. Should Sadiola be able to treat hard to use in an operating mine environment. Sequential Gaussian
sulfide material, mining could continue to the end of 2015 and simulation (or SGS) (Goovaerts, 1997), the industry norm at the
time, proved impractical because of the computing time involved
in generating the 50 realisations required per estimation zone.
1. Geology Manager, Geita Gold Mine, PO Box 532, Geita, Tanzania. The relatively new method of direct block simulation was
Email: [email protected] therefore selected.
S ROBINS
MAURITANI
Segala MALI
SENEGAL
Sabodala Loulou
Yalea
Tabakoto
Bamako
GUINEA
FIG 1 - Locality map for Sadiola mine, Mali, West Africa.
Dimitrakopoulos, Farelly and Godoy (2002), proposed the Senegalo-Malian Shear

direct block simulation (DBSIM) method as a viable alternative
to conventional simulation methods that considerably reduces
computational time. DBSIM detailed in Godoy (2003) and also
outlined in Boucher and Dimitrakopoulos (2009) and Peattie YATELA
(2005) simulates the internal points of each block, and when the SADIOLA
simulated block is calculated, the point values are discarded. The
simulated block value is then added to the conditioning dataset.
To integrate the block support conditioning data, the algorithm
has been developed in terms of a joint simulation, where the
second variable relates to the block values sequentially derived
through the simulation process. The algorithm simulates several SABODALA
LOULOU
hundreds of blocks per second and is considerably faster than YALEA
any point conditional simulation combined with reblocking. SEGALA
Furthermore, Godoy (2003), Benndorf and Dimitrakopoulos TABAKOTO
(2007) and others show that in addition to being substantially KENIEBA
faster and more efficient in terms of computing requirements, the
DBSIM method is reliable in terms of reproduction of the sample
statistics.
GENERAL GEOLOGY MEDINANDI

The Sadiola deposit is located within the Malian portion of the
Kenieba-Kedougou window, a major Early Proterozoic –
Birimian outlier along the NE margin of the Kenema – Man
shield (Figure 2). The Birimian of the window can be interpreted
as a collage of at least two NS trending terrains of different
nature. To the west, an older (±2.2 Ga) tholeitic mafic volcanic
unit with island arc type volcanics, intruded by a major
Keniebandi Saboussire Granitoids
calc-alkaline batholith, belongs to the Saboussire Formation. It is
separated from the dominantly sedimentary sequence of the Kofi Diale Dalema- Faults and
Formation by a major north-to-northeast trending shear zone. Kofi Shears
This sedimentary domain is significantly younger and is intruded
by calc-alkaline batholiths dated at 2.0 - 2.05 Ga. Within this FIG 2 - Geology of the Kenieba-Kedougou window.
USING GRADE UNCERTAINTY TO QUANTIFY RISK IN THE ULTIMATE PIT DESIGN
domain metamorphic grade is greenschist facies, with formation quartz-feldspar porphyry) and is spatially associated with a
of metamorphic biotite and locally amphibolite grade near major complex alteration pattern. Drilling of the (unweathered) primary
intrusions. The Kofi Formation is obliquely cut by the mineralisation has allowed detailed investigation of major and
approximately N-S to N10° trending Senegalo-Malian shear zone minor hydrothermal alteration processes that were active during
(SM), which is punctuated by several gold deposits along its the formation of the deposit. Alteration assemblages identified to
splays (Loulo, Yalea, Sadiola, Yatela), (Robins et al, 2005). date include calc-silicate, potassic, chlorite-calcite, carbonate
The Sadiola deposit is located in the north of the window and and silicification, and have allowed the deposit to be classified as
is hosted by sediments of the Kofi Formation, which have been a mesothermal-shear-zone hosted deposit. Gold is associated
intruded by numerous felsic intrusives. The dominant sediments with both arsenic and antimony dominated sulfide assemblages
consist of fine-grained greywacke and impure carbonates with including arsenopyrite, pyrrhotite, pyrite, stibnite and gudmuntite
minor tuffs and acid volcanics. At Sadiola, the intensely folded (Robins et al, 2005). Deposits of this type world-wide exhibit
impure carbonate packages comprise an alternation of limestone good continuity of mineralisation along the strike and extending
beds, a few millimetres to several decimetres thick, with thinner, to great depth. Structurally controlled, high grade pay shoots’
more detritic beds. The Sadiola deposit occurs along the N10° typically occur within a more pervasive lower grade halo. At
Sadiola, the location and geometry of high grade mineralisation
striking Sadiola Fracture Zone (SFZ), which is thought to be a
appears to be controlled by the confluence of the SFZ with the
brittle-ductile splay off the SM Shear at a sinistral releasing
N20° splays, resulting in steeply to vertically plunging zones
bend. The SFZ follows the steeply westerly dipping contact
within the plan of the SFZ. The geometry of the extensive soft
between greywacke to the west and impure carbonate to the east.
‘saprolitic’ oxide deposit at Sadiola relates almost exclusively to
Along the SFZ, both the greywacke and impure carbonate are the supergene weathering history of the primary mineralisation.
transposed. The SFZ and its wall-rock are injected by The permeability of the rock formation, controlled mainly by
discontinuous diorite dikes. Silicified quartz-feldspar-porphyry faulting, shearing and porosity, allows deep penetration of
(QFP) dikes often intrude along later steep west dipping, N20° ground water, causing oxidation of primary sulfides. Oxidation
striking structures. The QFPs cross-cut the diorites and may of pyrite (and other sulfide species) results in the formation of
show brittle fracturing. Post mineralisation deformation has sulfuric acid, further promoting the downward argillisation of the
complicated structural relationships in the deposit. bedrock to form the clay rich assemblages present in the
Gold mineralisation at Sadiola occurs along the SFZ over a saprolite. The irregular, ‘karst like’ soft rock/hard rock contacts
drilled strike length of approximately 2500 m, and remains open can be related to the extent of faulting and the original sulfide
to the north and south. N20° trending fault splays off the SFZ are content of the overlying profile. The intense weathering has
also well mineralised. The mine geological and grade block resulted in a tropical climate forming a series of decarbonated-
model indicates the presence of 20° to 25° south dipping ore argillised troughs of variable depth (up to 180 m depth) along the
shoots within the plane of the SFZ. Mineralisation occurs in all SFZ. This is the rich oxide orebody currently being mined
of the four major rock types (marble, greywacke, diorite, and (Figure 3).
FIG 3 - Section 5600N through Sadiola orebody. Mineralisation sits along the greywacke-marble contact (Sadiola Fracture Zone).
N20º faults are indicated in black. Section compiled by Dr A Smeesters, 2006.
S ROBINS
CREATION OF THE SIMULATION MODEL with the conditioning data zoned according to the estimation
domains (ZONECODE), outlined in Table 2. The input data
comprised raw grade information that was converted to normal
Simulation methodology space by the DBSIM program during the simulation runs. During
Although the deep sulfide project has focused specifically on the this process, DBSIM declustered the data, and output ‘equal
hard sulfide component of the Sadiola resource, a significant weighted’ statistics for the conditioning data, simulated nodes
amount of mineralised saprolite material occurs between the and blocks.
hard/soft boundary and the current LOM pit design. Since this
material would contribute towards the economics of the project, TABLE 2
it was necessary to account for it in the simulation model. The Table of estimation domains and their description.
laterite material only occurs close to the original ground surface,
and would have limited influence on the project economics since ZONECODE Description Project
most of the ore within this zone has already been mined. To 1000 Far north saprolite ore Main pit – oxide
speed up the simulation process, the laterite estimation domains
were excluded from the study. All other estimation domains used 2000 Saprolite waste Main pit – oxide
in the Sadiola recoverable resource model were honoured. 3000 North saprolite ore Main pit – oxide
Furthermore, the same densities as per the recoverable resource 4000 South saprolite ore Main pit – oxide
block model were retained in the simulation model (Table 1).
The determination of the various estimation domains in the 5000 Main ore – high grade Deep (hard) sulfide
recoverable resource model was based on trend analyses and 6000 Main Oore – low grade Deep (hard) sulfide
sample statistics. This process is outlined in Robins et al (2005). 7000 Footwall/hangingwall ore Deep (hard) sulfide
8000 North hard ore Deep (hard) sulfide
TABLE 1 9000 Waste Deep (hard) sulfide
Table of rock types and their associated densities.
ROCKTYPE Description Density Normal score semi-variogram models

code (g/cm3)
1 Laterite and clay 1.97 (average) Double structured, normal score spherical semi-variogram
models were calculated per ZONECODE (Table 3). The search
2 Saprolite 1.80 (average)
distances are those used for simulation in the x, y, and z
3 Silicified oxide 2.57 (average) orientations. The general characteristics of the normal score
4 Saprolite sulfide 2.00 (average) variograms can be summarised as follows:
5 Hard sulfide 1.55 to 3.03 (kriged) • The northern and southern mineralisation is separated by a
6 Blast oxide 2.10 (average) WNW trending fault that appears to have rotated the strike of
the mineralisation so that to the north, the mineralisation
7 Blast sulfide 2.10 (average) trends 030, and to the south, it trends 000.
• Most of the hard sulfide mineralisation (zones 5000, 6000
Sample data and 7000) is situated in the south of the orebody and
therefore their variogram structures resemble that of the
Almost all the deep sulfide sample data comprised NQ diamond South Saprolite ore (zone 4000), for which the relatively low
core at 1 m sample intervals. However, since a significant nugget is attributable to supergene enrichment of the grade.
amount of saprolite material – which was drilled with 125 mm The main ore high- and low-grade variograms (zones 5000
RC method and at 2 m sample intervals – would be included in and 6000 respectively) are similar in all respects.
the Deep Sulfide Project, all conditioning (sample) data was
composited to 2 m sample intervals. To test for clustering, data • The Far North (zone 1000) and North Saprolite (zone 3000)
was declustered using a moving window method. It was found ore zones have similar structure and orientation trending 030,
that there was no significant spatial clustering present in the data in line with the change in strike of the mineralisation. The Far
and therefore declustering was unnecessary. This was likely a North saprolite has significantly more continuity in the y
result of domaining the data and therefore achieving a relatively direction, though it was necessary to reduce the search
regular sample grid per domain. A sample input file was supplied distance in this direction to minimise negative kriging weights.
TABLE 3
Normal score semi-variogram models by ZONECODE.
ZONECODE C0 C1 C2 Range 1 Range 2 Search distance Azimuth

/dip
X Y Z X Y Z X Y Z
A2 A1 A3 A5 A4 A6
1000 0.190 0.510 0.300 14 35 15 50 130 30 30 30 20 030/00
2000 0.258 0.450 0.284 42 30 28 170 270 55 30 30 10 170/00
3000 0.200 0.400 0.398 10 10 12 40 50 30 25 40 15 030/00
4000 0.150 0.370 0.478 9 9 9 65 65 45 33 45 16 000/00
5000 0.300 0.370 0.328 11 11 11 35 35 23 65 65 30 000/00
6000 0.380 0.370 0.248 11 11 10 43 43 30 65 65 35 000/00
7000 0.350 0.270 0.479 9 9 8 33 33 25 120 120 60 000/00
8000 0.200 0.400 0.399 8 8 8 26 26 18 120 120 60 000/00
9000 0.231 0.301 0.467 13 13 13 132 132 105 60 60 30 000/00
The North Hard Ore (zone 8000) predominantly underlies the algorithm transformed them to normal score values for
North Saprolite, and has the same structure, though not the simulation, and then back-transformed the resultant normal score
same orientation. This is most likely a result of insufficient values to raw grade values. Both the raw and normal score data
sample data to define a trend. sets were output, though because they show the same trends, only
the raw data is presented below.
• All the hard sulfide variograms have isotropic horizontal
structures, attributable to the absence of close spaced grade
control data necessary to define the structures causing Statistical validation
anisotropy. The base statistics for the raw grade data are summarised in
• The waste variograms for the saprolite (zone 2000) and hard Table 4.
sulfide (zone 9000) are similar because of a relative lack of From Table 4, the base statistics indicated that the mean grades
grade variability between sample locations. This is to be and variances for the saprolite material (ZONECODES 1000
expected, considering that most of the mineralisation had to 4000) were generally slightly under-estimated with the
been included in the mineralised envelope. conditioning data, exhibiting values slightly higher than the
simulated values. Though the mean grades for the Saprolite
Representative number of realisations waste (ZONECODE 2000) and North Saprolite ore
(ZONECODE 3000) were only marginally lower, both the mean
To determine the number of realisations that would be required to
grade and variance for the South Saprolite ore (ZONECODE
obtain a reliable uncertainty model, the number of realisations
were plotted against the progressive mean and progressive 4000) were significantly lower than the corresponding
Coefficient of Variation (COV) for each simulated domain. The conditional data values. This was initially a cause for concern for
number of simulated blocks for each domain was greater than the pit optimisation phase of the project because the South
95 per cent of the total number of available blocks. For the first Saprolite ore comprises the majority of the saprolite material
few realisations, the mean and COV values in the plots ‘bounced below the current LOM pit. However, the total saprolite ore only
around’, but as the progressive values included more realisations, comprises 16 per cent of the total ore tonnes between the current
they stabilised, and where they levelled sufficient realisations, they LOM and Deep Sulfide pit designs. The far north Saprolite
were considered to have been used to model the variability. In this (ZONECODE 1000) is relatively insignificant regarding its ore
study, the grade for five out of the nine simulation domains tonnage contribution towards the Deep Sulfide Project.
could have been considered sufficiently simulated with up to From the base statistics, the simulation mean and variance for
25 realisations, however, the remaining four domains required the main ore high grade and low grade (ZONECODEs 5000 and
additional realisations. Since it was easier to deal with 6000 respectively) – the primary drivers for the Deep Sulfide
50 realisations than 25 realisations in the probability calculations, Project – though marginally high, are reasonably close to the
and most of the Deep Sulfide ‘ore’ zones required more than mean and variance of the conditioning data, which fall within
25 realisations, it was decided to use 50 realisations for all the the range of values of the ten realisations. Similarly for the
mineralised domains, and 25 realisations for the ‘waste’ domains. Hangingwall and Footwall mineralisation (ZONECODE 7000),
Thus the 25 realisations for ZONECODE 2000 and 9000 were the conditioning data mean and variance are within acceptable
repeated for the second 25 (26 to 50) mineralised domain limits of the corresponding ten realisation values. The
realisations. In this way, the simulation time was reduced by not simulated mean and variance values for the North Hard
performing unnecessary simulations on the ‘waste’ domains, while ore (ZONECODE 8000) are significantly higher than the
simultaneously achieving sufficient variability in the ‘ore’ corresponding conditioning data. A poor result, however, is to
domains, which were relatively quick to simulate. be expected since this domain has been insufficiently sampled.
Furthermore, the North Hard ore comprises a relatively
Initial validations on point data insignificant proportion of the total Deep Sulfide ore and
The internal nodes were estimated on 2.0 m × 2.0 m × 1.25 m therefore is not expected to have significant influence on
centres for each simulation, totalling 100 nodes per regularised the pit optimisation runs. For the Hard Sulfide waste zone
(SMU) block. These were only output for the first ten (ZONECODE 9000), the simulated mean grade (0.50 g/t to
simulations in order to check that the results honoured the 0.54 g/t) was significantly higher than the mean of the
conditioning data, the histogram of the normal score and raw conditioning data (0.26 g/t), though the variance is similar. This
data, and the variograms of the normal score and raw data. For zone was expected to remain largely subeconomic and was
the remaining simulations, only the block values were retained. therefore unlikely to impact significantly on the pit
Raw grade values were input into the program. The DBSIM optimisation runs.
TABLE 4
Summary base statistics for raw grade point data.
ZONECODE Conditioning data Simulation range in mean value Simulation range in variance
Mean Variance Minimum Maximum Minimum Maximum
1000 1.65 7.27 1.44 1.62 4.87 6.55
2000 0.36 1.69 0.31 0.33 0.72 1.21
3000 2.03 25.74 1.86 2.02 17.01 23.36
4000 2.93 32.18 2.62 2.71 19.99 25.13
5000 2.84 8.73 2.82 3.03 8.84 16.43
6000 1.36 6.26 1.35 1.44 5.68 7.27
7000 1.14 6.94 1.17 1.33 6.24 10.65
8000 2.09 28.49 2.16 2.54 19.30 47.66
9000 0.26 0.54 0.50 0.54 0.54 0.73
S ROBINS
Histogram validation
Histograms were generated for the first ten realisations of both
the raw grade and normal score simulation nodes. Since the two
data sets are similar, only the raw data is discussed. The raw
grade histograms for the saprolite material (ZONECODES 1000
to 4000) indicated that, except for the North Saprolite ore
(ZONECODE 3000), the simulated grade for the saprolite
material has been understated. The Saprolite material
(ZONECODES 1000 and 4000) was only marginally so, but the
grade for the Saprolite waste (ZONECODE 2000) has been
significantly understated. Since this is a ‘waste’ domain, with
only five per cent of the conditioning data above an economic
cut-off grade to begin with, this grade understatement is not
expected to have a significant impact on the pit optimisation
runs. For the primary drivers of the Deep Sulfide Project, the
Main Ore High Grade and Low Grade (ZONECODES 5000 and
6000 respectively), the raw grade histograms show good
correlation between the ten realisations and the conditioning
data. Figure 4 illustrates this deduction for the main high-grade
mineralisation (ZONECODE 5000).
Cumulative Frequency Distribution - ZONECODE 5000 Raw Grade Data
105%
100%
95%
90%
85%
Cum ulative Frequency (%)
80%
75%
70%
65% FIG 5 - Normal score variogram model for ZONECODE 5000.
60%
55%
50%
45%
40% model for the Main High-Grade mineralisation (ZONECODE
35% 5000). This was generally the case for the other zones,
30%
25% particularly the Hard Sulfide mineralised domains
20% Realisations 1 - 10
15% Conditioning Data
(ZONECODES 5000 to 8000).
10%
0 2 4 6 8 10 12 14 16 18 20 22 24 26 28 30
Upper Class Limit (g/t) Post simulation processes
After the simulation process, the base statistics per block for all
FIG 4 - Cumulative frequency distribution for ZONECODE 5000 the realisations were calculated. The following statistics were
raw grade data. calculated for each block:
For the Hard Sulfide Hangingwall and Footwall mineralisation

(ZONECODE 7000), the simulated grade appears to be mean, m =
∑ AU (1 to n)
marginally overstated between 3 g/t and 5 g/t. The same is n

evident for the North Hard ore (ZONECODE 8000) between
5 g/t and 7 g/t. Both of these domains require additional drilling where:
to ensure they are representatively sampled. The histograms for n is the number of realisations
the Hard Sulfide ‘waste’ domain (ZONECODE 9000) show that
the grade is significantly overstated up to about 2.5 g/t, however
94 per cent of the conditioning data is below the lower marginal variance, s2 =
∑ (AU − m) 2
cut-off grade of 1.20 g/t. The overstatement of grade is therefore (n − 1)

unlikely to have a significant impact on the pit optimisations.
From a statistical perspective, the Sadiola simulation model can standard deviation, s = s2
be considered to be representative of the conditioning data.
s
Variogram validation coefficient of variation, COV =
m
The final test to determine the validity of the simulation model
was to test the reproduction of the spatial correlation between the Similarly, the probability of each block grade being above 2.0 g/t
simulated results and the conditioning data. Experimental was calculated.
variograms were calculated for the first ten realisations and
plotted against the semi-variogram model that was input into the No of realisations > 2.0g/t
simulation process. To facilitate this process, the DBSIM probability, P(AU > 2.0 g/t) =
program converted the input semi-variogram models to block Total no of realisations (n)
support and output the results. The resultant normal score
variogram model for the Main High-Grade mineralisation (zone Next, for pit optimisation purposes, the laterite domains were
5000) is presented in Figure 5. added to the simulation model, with each realisation adopting the
Although not an exact match, the normal score simulated estimated grade for the laterite model blocks. The final simulation
experimental data compares well with the regularised variogram model was depleted by the saprolite life of mine volume.
The final simulation model contained primary fields for coefficient of variation provides a reliable indication of uncertainty
50 grade realisations, coefficient of variation (providing an in the simulated grade, it can also be used as a tool for confirming,
indication of uncertainty in the grade value), and probability that or even improving, the classification of the Deep Sulfide
the grade would be greater than 2.0 g/t. Secondary fields were mineralisation. Furthermore it has potential to direct drilling
the remaining base statistics, the initial classification and rock programs in zones that require additional sample information, for
type fields, and the ZONECODE field. A representative section example, the Footwall zone to the east.
(EW-5550) is presented to illustrate the primary fields of the
final simulation block model (Figure 6 to Figure 8). The $475/oz PIT OPTIMISATIONS AND DATA PROCESSING
prefeasibility pit shell is shown in white, and the models have
been depleted to the current ‘oxide’ LOM design. For each simulated orebody model, an ultimate pit shell, or
From Figures 6 and 7 it is evident that although realisation one maximum cash flow pit, was derived using the Lerch-Grossman
shows significant material within the prefeasibility pit shell algorithm (Lerchs and Grossmann, 1965). The algorithm
above 2.0 g/t, the probability of the actual grade being above essentially begins at the model surface and ‘mines down’ into the
2.0 g/t is low. For this reason, one shouldn’t place too much model making two basic decisions for each block:
emphasis on the grade of a single realisation, but rather consider 1. Should the block be mined?
a number of realisations simultaneously. The use of probability
calculations provides a good overview of the results of all the 2. Does the material go to the treatment plant or the waste
realisations simultaneously. dump?
Figure 8 illustrates how the coefficient of variation calculation These two decisions are essentially based on the revenue
provides an indication of uncertainty in the simulated block grade. generated by the block at a defined gold price, which is offset
The diagonal white lines are borehole traces, and it is evident that against the cost of mining the block and the cost of processing
zones of high confidence (low uncertainty) correspond well with the material. Although valued individually, the decision to mine a
the borehole traces, while zones of high uncertainty correspond block or not is determined by considering the surrounding blocks
with zones that have been insufficiently sampled. Since the in conjunction with the block in question. For example, mining a
W E
FIG 6 - Simulated block grade for realisation 1 on Section EW-5550.
W E
FIG 7 - Probability of grade greater than 2.0 g/t on Section EW-5550.
S ROBINS
W E
FIG 8 - Block coefficient of variation on Section EW-5550.
high grade block below waste blocks may in fact pay for the cost Calculation of probability/risk pits
of mining the waste blocks and still generate a profit. The
ultimate pit is thus generated from all the blocks that the The probability of mining each block was determined from the
optimiser decides to mine. The pit optimisation parameters were number of times it was mined in each of the 50 ultimate pits.
used to optimise the simulation realisations (Tables 5 and 6). Each block was then flagged in the simulation model, based on
The 50 NPV results for the various realisation ultimate pits its probability of being mined, and risk pit shells were generated
were plotted with the NPV for the recoverable resource model at five per cent probability intervals from five per cent to 95 per
(Figure 9). From Figure 9 it is evident that the NPVs for all the cent (Figure 10 and Figure 11).
simulation realisations are greater than the NPV for the The $475/oz prefeasibility pit shell is shown as a white line in
recoverable resource model and that most of them are Figures 10 and 11. It is evident in both figures that the
significantly greater. This indicates that significant upside prefeasibility pit does not fetch a significant amount of high
potential exists for the Deep Sulfide Project. Although this is probability/low risk material, and that in the east (Figure 10), it
indicative of upside potential, it would be more useful to take this mines low probability/high risk Footwall material. The Footwall
information a step further and calculate probability or risk pits is known to be under-drilled and is the target for the Phase 8
from the individual realisation pits. Deep Sulfide drilling program.
TABLE 5
Pit optimisation parameters.
Gold price $475

Selling cost 7% (6% royalty + 1% management fee)
Annual throughput 5.2 Mt
Rock type Mining Processing

Mining unit cost Mining CAF Processing unit cost Recovery
US$/t US$/t
Laterite (ox) 1.70 1.02 17.05 0.93
Saprolite oxide – soft 1.67 1.00 16.30 0.93
Siliceous oxide – hard 1.58 0.94 22.67 0.93
Saprolite sulfide – soft 1.50 0.90 21.01 0.8
Hard sulfides – hard 1.32 0.79 20.29 0.8
Intermediate oxide 1.59 0.95 17.05 0.93
Intermediate sulfide 1.59 0.95 21.76 0.8
Mixed oxide 1.59 0.95 22.67 0.93
Mixed sulfide 1.59 0.95 21.52 0.75
$/tonne/bench (10 m)
Applicable to 160 level
Adjustments 0.0159
TABLE 6 3.93 g/t), though the tonnes above cut-off show significant
upside potential.
Slope parameters for the pit optimisation.
• P80: The grade tonnage curve is similar to that of P50 at
North/south division Slope angle Domain Slope 2.0 g/t cut-off, with the SMU grade of 3.79 g/t falling in the
0 to 180 35 1 East oxides middle of the realisation grade distribution of 3.67 g/t to
180 to 359 40 1 West oxides 3.93 g/t. Upside tonnage potential is indicated above cut-off.
0 to 180 52 2 East sulfides • P95: Similarly indicates significant upside potential to the
tonnage above 2.0 g/t, with the SMU grade of 3.83 g/t
180 to 359 48 2 West sulfides
midway between the realisation grade limits of 3.74 g/t and
4.0 g/t.
Generation of grade tonnage curves Except for P05 – for which the average grade above 2.0 g/t
cut-off is marginally low - the above probability pit shells indicate
For each of the 50 realisations, and the recoverable resource that the SMU grade has been estimated with reasonable accuracy,
(SMU) model, grade tonnage curves were generated for the P05, even though within all the shells, the tonnes above cut-off have
P50, P80, P95, and prefeasibility probability pit shells. Less been understated in the SMU model. The tonnage and grade
variability than expected was achieved in each of the grade results are not too dissimilar despite the different techniques.
tonnage curves:
Within the prefeasibility pit shell, a comparison of the simulation
• P05: At the approximate economic cut-off grade of 2.0 g/t, and SMU model grade tonnage curves indicates that at 2.0 g/t
the SMU model grade was 3.70 g/t, while the simulated cut-off, both the SMU grade and tonnage curves lay within the
realisation grades varied between 3.65 g/t and 3.87 g/t. The range of the realisation distribution. This is illustrated in
realisations indicated significant upside potential for both Figure 12, where the tonnage curve has been re-scaled to a
grade and tonnage above cut-off. maximum of 70 million tonnes. Although one may argue that the
• P50: At 2.0 g/t cut-off, the SMU grade of 3.76 g/t falls in the SMU model is marginally too selective, Figure 12 indicates that
middle of the realisation grade distribution (3.65 g/t to this selectivity is acceptable. Above 2.0 g/t cut-off grade, the
NPV - Simulation Model Pits vs Recoverable Resource Model Pits

220
200
180
160
NPV (US$ Million)
140
120
100
80
60
40
20
-
1
3
5
7
9
11
13
15
17
19
21
23
25
27
29
31
33
35
37
39
41
43
45
47
49
Simulated Realisations Recoverable Resource Realisation Number
FIG 9 - NPV for simulation realisations and recoverable resource model.
FIG 10 - Plan view (elevation -90 m) showing the probability of mining.
S ROBINS
W E
FIG 11 - Section EW-5550 showing probability of mining.
Grade Tonnage curves for the Prefeasibility Pit Shell TABLE 7

70 7
Ore/waste cut-off grades.
60 6
Average Grade Above Cut-
Rock type Cut-off (g/t) Rock type Cut-off (g/t)

50 5
Laterite 1.10 Hard sulfide 1.90
Tonnes (Millions)
off (g/t)
40 4
Saprolite 1.00 Intermediate oxide 1.10
30 3
Siliceous oxide 1.50 Intermediate sulfide 1.60
20 2
Saprolite sulfide 1.60
10 1
0 0 The calculated open pit values using the recoverable resource

0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0 model for each of the probability pits (P05 to P95 at 0.5 intervals)
Cut-off (g/t)
are presented below in Figure 13. As expected, an inverse
SMU Tonnes SMU Grade
relationship exists between the pit value and stripping ratio with
increasing mining volume. There is an interesting kink in both
FIG 12 - Grade tonnage curves within the prefeasibility pit shell. curves for probability pit P80, which leads to its inclusion in
further pit value analyses. These additional analyses involved
comparing the values for each of the simulation model realisations
average grade above cut-off for the SMU model is 3.82 g/t, while with the SMU model value within selected probability pits. Open
the realisation grade ranges from 3.75 g/t to 3.99 g/t. At the same pit values were calculated for selected probability/risk pits (P05,
cut-off grade, the SMU tonnage of 19.5 Mt falls within the P50, P80, P95) and the $475/oz prefeasibility pit using all
realisation distribution range of 19.0 Mt to 21.0 Mt. 50 simulated realisations and the recoverable resource (SMU)
model. Table 9 summarises the pit value base statistics for the
Valuation of the pit shells 50 realisations and SMU model per pit shell, and Figure 14
illustrates the distribution of the realisation pit values.
A pit value was calculated for each of the probability pits (P05 to
P95 at 0.5 intervals) using the recoverable resource model. Since Probability Shell vs Open Pit Value
the extraction of the resource within these shells was not 50 10.6
scheduled, it was not possible to calculate an NPV for each pit, 40 10.4
30 10.2
Open Pit Value (US$) - (Millions)
but an open pit value was obtained for each scenario, based on 20 10.0
the cost of mining ore and waste, the cost of treating the different 10 9.8
Stripping Ratio
material types, and the revenue generated at a gold price of 0 9.6

$475/oz. The 2006 grade control cut-off grades were used to -10 9.4
-20 9.2
define ore and waste material for the various rock types
-30 9.0
(Table 7), with material above the cut-off categorised as ore. -40 8.8
The cut-off grades correspond with ‘break-even’ marginal -50 8.6
-60 Prob vs Val 8.4
grade material. The various parameters of recovery, cost, and Prob vs SR
-70 8.2
revenue that were used in the pit valuation calculations are -80 8.0
presented in Table 8. The cost values are those calculated by the 5 10 15 20 25 30 35 40 45 50 55 60 65 70 75 80 85 90 95
Sadiola long-term mine planner in February 2006. The Probability Shell / Risk Pit
rehabilitation (rehab) costs had been included in the variable

process costs and were therefore set to zero. FIG 13 - Open pit values for the probability pits.
TABLE 8 Histogram of Open Pit Profit Values

Cost parameters applied during pit valuation. 22
20
Item Material type Value
18
Mine call factor 100%
16
Recovery saprolite oxide 93%
14
Frequency
Recovery saprolite sulfide 80%
12
Recovery hard sulfide 80%
10
Fixed mining ($/tonne treated) 2.40
8
Fixed process costs ($/tonne) 1.07 6
Laterite (oxide) 9.84 4
Saprolite oxide – soft 9.09 2
Siliceous oxide – hard 15.46 0
Variable process costs ($/tonne) Saprolite sulfide – soft 13.80 -20 0 20 40 60 80 100 120 140 160 180
Hard sulfide – hard 15.58 Class Mid Point - Profit ($ Million)
Intermediate oxide 9.84 P05 P50 P80 P95 PPF
Intermediate sulfide 14.56

FIG 14 - Histogram of the open pit values for the 50 realisations.
Admin costs ($/tonne) 3.59
Capital replacement ($/tonne) 0.46
Pit Values for Simulations and SMU Model
Rehab costs ($/tonne) 0.00 200
Laterite (oxide) 2.089
150
Saprolite oxide – soft 2.058 Pit Value ($ Millions)
100
Siliceous oxide – hard 1.938
Variable ore ($/tonne) Saprolite sulfide – soft 1.852 50
Hard sulfide – hard 1.825 0
Intermediate oxide 1.960 -50
Intermediate sulfide 1.960 -100

P05 P50 P80 P95 PPF
Laterite (oxide) 2.083
Probability Pit
Saprolite oxide – soft 2.051 Realisation (1 to 50) SMU Model
Siliceous oxide – hard 1.933
Variable Waste ($/tonne) Saprolite sulfide – soft 1.846 FIG 15 - Pit value per probability pit for the simulation and SMU
models.
Hard sulfide – hard 1.820
Intermediate oxide 1.954
confirms that the prefeasibility pit is the optimal pit for the Deep
Intermediate sulfide 1.954 Sulfide project, given the $475/oz gold price and cost
Gold price ($/oz) 475 parameters. The SMU model produces a profit of $112 million
and falls in the middle of the distribution of the simulated
realisations. The risk does exist that a profit of only $64 million
TABLE 9 may be realised, however a profit of $179 million is also
Base statistics for the realisation pit values per pit shell. possible, with the most likely profit ranging between $87 million
and $148 million. It is evident from Figure 15 that the profit
Pit SMU Mean Std dev Max profit Min profit curves flatten off significantly from the P50 to the P95 risk pit,
profit profit ($ million) ($ million) ($ million) and that there is no insignificant difference in their mean profit
($ million) ($ million) (Table 9). Though a marginal decision at a gold price of $475/oz,
P05 -78.010 -2.621 34.782 86.379 -67.288 mining an equivalent of the P80 or P50 pits could dramatically
P50 14.011 85.385 31.119 160.421 30.655
improve the Deep Sulfide reserve ounces, and the P50 pit would
most probably become a real option should the gold price
P80 21.745 82.204 24.921 139.124 37.691 increase significantly.
P95 42.085 89.025 21.389 136.528 47.883
P-PF 112.889 108.036 24.400 179.400 60.870 CONCLUSION
The final risk pits generated from the simulation model indicated
From Figure 14 and Table 9 it is evident that the $475/oz that the prefeasibility pit is generally conservative, corresponding
prefeasibility pit offers the greatest profit margin but not the least with the 100 per cent probability shell. There is however a
risk. Probability pit P95, not surprisingly, has the tightest certain amount of risk involved with mining the Footwall
distribution about the mean and would therefore offer the least mineralisation, where the prefeasibility pit digs on
risk to achieving the indicated profit. However, considering the high-risk/low-probability material. This is confirmed by the COV
respective mean profit values for the prefeasibility and P95 pits, values in this zone, which indicate high uncertainty in the grade
the value offered by P95 is significantly lower than that of the estimate. Based on these results, an infill drilling program has
prefeasibility pit. Given the relatively minor additional risk, the been planned to mitigate this risk. At the current economic
prefeasibility pit may be the preferred pit option. Figure 15 cut-off grade of 2.0 g/t, grade tonnage curves generated for
S ROBINS
selected risk pits have shown that although upside tonnage REFERENCES
potential exists for the Deep Sulfide Project, the recoverable
Benndorf, J and Dimitrakopoulos, R, 2007. New efficient methods for
resource (SMU) model grade estimate is reliable. With the conditional simulation of large orebodies, in Orebody Modelling and
current cost information, and a gold price of $475/oz, the Strategic Mine Planning, second edition (ed: R Dimitrakopoulos),
prefeasibility pit is optimal. The calculation of economic pit pp 103-110 (The Australasian Institute of Mining and Metallurgy:
values for the realisations within the same selected risk pits Melbourne).
indicated that the P50 to P95 pits and the prefeasibility pit are Boucher, A and Dimitrakopoulos, R, 2009. Block-support simulation of
not significantly different. Therefore, there is potential to add multiple correlated variables, Mathematical Geosciences,
reserve ounces to the project without significantly decreasing its 41(2):215-237.
value. An increase in gold price, or a decrease in operating costs, Dimitrakopoulos, R, Farelly, C T and Godoy, M, 2002. Moving forward
would make this additional material available to the project. At a from traditional optimisation: Grade uncertainty and risk effects in
open-pit design, Transactions of the Institutions of Mining and
gold price of $475/oz, and with current cost parameters, the Metallurgy, Mining Technology, 111:A82-A88.
prefeasibility pit shows maximum value for the project with no Dohm, C, 2005. Mineral resource classification from innocence to
significant additional risk. excellence and beyond, presentation made to the Anglo American
This study has highlighted the potential to use the COV values Mine Geology Conference 2005, MinRED, 24 p.
as a reliable classification method. It is recommended that the Godoy, M, 2003. The effective management of geological risk in
classification is re-run using the simulation COV information, as long-term production scheduling of open pit mines, PhD thesis
proposed by Dohm (2005). This exercise would eliminate (unpublished), University of Queensland, Brisbane.
unnecessary drilling and potentially save drilling costs. The gold Goovaerts, P, 1997. Geostatistics for Natural Resources Evaluation,
price used for this project was $475/oz (the recommended pp 370-382 (Oxford University Press: New York).
Anglogold Ashanti resource reporting gold price for 2005). Lerchs, H and Grossmann, L, 1965. Optimum Design of Open-Pit Mines,
Trans CIM, LXVII:17-24.
Anglogold Ashanti has since increased their resource reporting
Peattie, R, 2005. DBSIM – Direct Block Simulation Algorithm User’s
gold price to $1000/oz, and at $720/oz, their reserve reporting Manual (Draft 1), unpublished document, AngloGold Ashanti, South
gold price is significantly greater than the gold price used for this Africa, 18 p.
project. Since the current gold price provides significant upside Robins, S P, Nicholls, C E, Fouche, F and De Hert, G, 2005. Sadiola –
potential for the Deep Sulfide Project, it is recommended that the Identified Mineral Resource Statement 2005 – Part Two A – Sadiola
pit shell work is rerun using a suggested gold price of $700/oz Main Pit and Hard Sulfides, unpublished company report, SEMOS,
and the latest mining and processing costs derived from the SA, Mali, pp 38-39.
prefeasibility project. Van der Westhuizen, R, 2005. Sadiola Reserve Report December 2005,
unpublished company report, SEMOS, SA, Mali.
HOME
Angle Reverse Circulation Versus Blastholes in Ore Control –

Sampling Quality Versus Data Density – A Case Study
I Douglas1, B Perry2, W Hardtke2 and L Allen3
ABSTRACT Other issues were of concern in choosing the best approach to

ore control, such as obtaining good samples from blastholes and
Many open pit mines face a difficult choice between the cost of special
drilling programs for better sampling of ore control and sampling the the possible presence of water in the mining areas. These studies
production drilling, usually with blastholes. This choice is not always were completed to help guide the decision regarding which
made on the basis of detailed studies, but on past experience and any approach to take in primary ores. The ore zones were noted to be
biases that those charged with making the decision may have. This can continuous between saprolite and primary rock. Initial
lead to significant costs, either in the execution of unnecessary drilling reconciliation in the oxidised ores was poor, with ore control
programs or poor definition of ore and waste – sometimes both. This finding more tonnes at lower grade than predicted by the
paper outlines a methodology used by Newmont Mining to complete a exploration model. Blasthole sampling was introduced after nine
cost-benefit study and present results of the study from a theoretical months of continued poor reconciliation within primary rock,
viewpoint, as well as from production. The approach included generating with the site making a full change to blasthole ore control in
conditional simulations, extraction of various vertical blasthole and angle
March 2007. Improvement in reconciliation was measurable after
reverse circulation (RC) data sets representing potential configurations,
completing ore control utilising different methodologies and tabulation three months.
and comparison of results. The study concluded that blasthole sampling
was the preferred approach for this deposit. The conclusion was a LOCATION AND GEOLOGIC SETTING
function of:
• a planned blasthole spacing that is quite close, The Ahafo gold property is located in western Ghana, between
• the close spaced data allowing the information of surrounding the towns of Kenyase and Adroboa in the Brong-Ahafo Region,
blasthole assays to correct for local sampling variances coming from about 290 km northwest of Accra. It is 107 km northwest of
poorer quality sampling, and Kumasi, and 40 km south of the regional capital of Sunyani
• the prohibitive cost of close spaced angle RC spacing needed to (Figure 1). The project is composed of a southern and northern
correctly classify ore and waste. area, with the plant site and current operations located in the
It is acknowledged that this would likely not be the conclusion if southern area. The Apensu deposit that is the subject of the
blasthole sampling were significantly biased. Actual production data has conditional simulation study (Figure 2) is also located in the
demonstrated that systematic biases are not present in the blasthole southern area, adjacent to the plant site. The Ahafo project is
samples. Mill reconciliation to production demonstrates that the polygons located in the West African Craton, within the Palaeoproterozoic
are well defined, and excessive dilution from misclassification is not
Man Shield. The project is located within the Shield, in the
occurring.
eastern terrain underlain by Palaeoproterozoic Birimian rocks.
The Birimian rocks consist of five evenly-spaced tholeiitic to
INTRODUCTION acidic composition volcanic belts trending northeast-southwest.
Newmont acquired the Ahafo project through its merger with Basins between the volcanic belts are filled by predominantly
Normandy Mining in 2002. Following a successful feasibility turbiditic sediments and the transition zones between the
study in 2003 and construction in 2004 - 2006, Newmont volcanic rocks and the sedimentary rocks are filled with
commenced mining operations from two open pits, Apensu and chemical sediments. Three granite successions have intruded the
Subika, in Ahafo in January 2006, with plant wet commissioning Birimian rocks. The Ahafo project is located in the Sefwi Belt,
in June 2006. During 2007, mining products consisted of 8 Mt of one of the five Birimian volcanic belts. The Sefwi Belt is
oxide, transitional, and primary ore, grading 1.88 g/t Au and an bounded to the northwest by the Sunyani Basin, and to the
associated 38 Mt of waste. In 2005, finalisation of the ore control southeast by the Kumasi Basin. Dixcove-type or ‘belt-type’
system and approach were under review. The mining approach granitoids intrude the contact and predominate in the
selected was to drill and blast on 9 m benches and mine on 3 m metavolcanics that form more or less elongate bodies parallel to
flitches. Two options were considered for ore control approach in the regional strike. Volcanic rocks in the belt are mainly basaltic
primary ores: and have been metamorphosed to varying degrees from lower
1. drill three 9 m benches with angle reverse circulation (RC) greenschist to lower amphibolite facies. The sedimentary
sampling on 2 m intervals, or succession consists mainly of fine to medium-grained lithologies
(argillites and wackes) with variable amounts of volcaniclastic
2. sample the blastholes (9 m) and split into three 3 m samples
material. Cape Coast type granitoids intrude the metasedimentary
for assaying.
rocks. The strong north-easterly grain of the Palaeoproterozoic
Because no blasthole drilling was completed in oxidised terrane parallel to the volcanic belts is the predominant trend that
saprolite ores which were free-digging, a full set of angle RC governs major structural features. The faults and associated
data was completed in oxidised rock, which formed the structures display a complex history of movement, including thrust
conditioning data for simulation studies completed at Apensu. faulting and shearing with both normal and strike-slip motion and
have played a major role in emplacement of mesothermal gold
mineralisation. Regional structure is controlled by the Kenyase
1. MAusIMM, Newmont Mining Corporation, 10101 E Dry Creek Road,
Englewood CO 80112, USA. Email: [email protected]
Thrust Fault, a northeast to southwest trending regional thrust that
separates the Sefwi Belt from the Sunyani Basin.
2. Newmont Mining Corporation, 10101 E Dry Creek Road, Englewood Three discrete geological zones are recognised within the
CO 80112, USA.
Ahafo project, relating to currently identified mineralisation
3. MAusIMM, Newmont Mining Corporation, 10101 E Dry Creek Road, styles. The Ahafo district currently consists of 12 discrete
Englewood CO 80112, USA. Email: [email protected] deposits, localised along multiple northeast-striking structural
I DOUGLAS et al
FIG 1 - Project location plan.
zones. All of the ‘shear zone’ type deposits appear to be part of 3. mylonites after granitoid of weak to strong foliation and
the same mineralised system. As with many deposits located in local folding within the Kenyase Thrust Fault;
tropical climates, a saprolite zone, typically between five and
50 m thick, is developed at the surface. The saprolite zone gives 4. multiple thrust fault duplexes consisting of mixed mylonite
way at depth to a sulfide zone, where gold occurs in structurally after granitoid and volcano-sedimentary units, containing
controlled zones of hydrothermal alteration. Six lithologies and/or rigid granitoid fragments in a matrix of plastically-deformed,
litho-structural units have been recognised in the Kenyase-style predominantly chloritic and volcano-sedimentary units;
zones, consisting of:
5. phyllonite units, dominated by phyllosilicate minerals,
1. weakly foliated mixed (meta)-pelitic sedimentary rocks and resulting from plastic deformation of the volcano-
(meta)-mafic volcanic units in the footwall of the Kenyase sedimentary units; and
Thrust Fault;
6. tectonic breccias (cataclasite) and finely-crushed rock
2. Dixcove Suite granitoids form the hanging wall to the formed within the granitoid units by local, brittle reactivation
thrust; on or near the Kenyase Thrust Fault.
ANGLE REVERSE CIRCULATION VERSUS BLASTHOLES IN ORE CONTROL – SAMPLING QUALITY VERSUS DATA DENSITY
FIG 2 - General layout, Ahafo project.
Hydrothermal alteration in Kenyase Style deposits is 0.6 g/t, was irregular, somewhat irregular on section and not very
characterised by the alteration of primary chlorite to sericite-silica, predictable from section to section. The only available tool to
and the addition of silica, iron carbonate, pyrite and local albite to provide predictive capability for making decisions on how to
the host rock. Ore zones, while structurally controlled, do not have approach ore control was considered to be conditional
visual identifiers that can be used reliably to determine ore and simulation. An area of approximately 300 m (along strike) by
waste. Discontinuities are abundant at the ore control scale, 150 m was drilled on approximately a 12.5 m × 12.5 m grid.
whereas overall geometry is reasonably continuous. An example Holes were angled at 60° into the Kenyase thrust, approximately
cross section of the Apensu deposit is depicted in Figure 3. perpendicular to the mineralisation (Figures 4a and 4b).
A set of ten simulations using the sequential Gaussian
CONDITIONAL SIMULATION FRAMEWORK simulation method was completed over the test area on 1 m × 1 m
horizontal node spacing and 3 m vertical spacing. The program
In preparation for initial mining in January 2006, an angle RC SGSim of GSLIB (Deutsch and Journel, 1998) was used. After
data set was available for use in generating conditional reviewing statistics and comparing them with the original data, a
simulations for determining future ore control options. Initial single simulation was selected as the ‘ground truth’ based on the
review of the angle RC data set showed that while the main zone closest visual and statistical comparison to the original RC data.
was continuous, the ore-waste boundary, at approximately Once the base simulation was selected, vertical blasthole and
I DOUGLAS et al
FIG 3 - Cross-section through the Apensu deposit.
FIG 5 - Extracted data sets and source simulation. (A) Blasthole

data set, (B) source simulation and (C) angle RC data set.
Recognising that sampling of blastholes has numerous

challenges that usually result in less than an optimal sample being
taken, a random error component was tested on the blasthole data
set before its use in the ore control tests. A random error
FIG 4 - Source data for conditional simulation. (A) Oblique view of component was added to the blasthole data set. This random error
Angle RC and raw assays. (B) Cross-section displaying 3 m was tested in ranges from ±10 to ±90 per cent. Errors from the ±10
composited grades and outline at 0.6 g/t. to the ±50 per cent tests showed little difference in results, and at
±90 per cent, the difference was approximately two per cent on
angle RC data sets were extracted from the simulation (Figures 5a, contained ounces inside ore control polygons using the affected
5b and 5c) and used in subsequent testing of ore control. The data set. For the purposes of the study, a subjective choice of
blasthole data was extracted at a 6 m × 6 m spacing, and the ±20 per cent random error was applied to the extracted blasthole
angle RC on a 12.5 × 12.5 m spacing. This allowed 36 different data set in all consequent tests. Ore control comparing the
blasthole data sets by moving the origin 1 m each time and 25 blasthole data sets to the angle RC data sets was completed both in
different angle RC data sets by moving 2.5 m each time. an automated way as well as using a manual definition.
ANGLE REVERSE CIRCULATION VERSUS BLASTHOLES IN ORE CONTROL – SAMPLING QUALITY VERSUS DATA DENSITY
Automated ore control test Automated results

The automated approach removed the subjectivity of defining The polygon outlines derived from the SMU blocks at 0.6 g/t
polygons by hand and assumed that the polygons defined from were used to filter the conditional simulation into nodes routed as
SMU block outlines would be mineable. This approach was ore and waste. By tabulating the nodes inside the polygons
applied to both drill types (blasthole and angle RC). These (routed as ore), one could compare the results expected from the
polygons were not directly compared to hand drawn polygons. two approaches with the ‘actual’ results. These were tabulated
Both the blasthole data based ore control polygons and the angle for all 36 possible outcomes from the blasthole and 25 possible
RC based ore control polygons were generated by ordinary kriging outcomes from the angle RC data. The tabulations were
to an SMU-sized block of 5 × 5 m horizontal and 3 m vertical. For completed on a net revenue basis, with process cost, recovery
the blasthole set, 36 sets of polygons were generated from the 36 and drilling cost included in the revenues. Because the blastholes
possible blasthole data sets. For the angle RC set, 25 sets of will be drilled irrespective of the chosen method, the only cost
polygons were generated from the 25 possible angle RC data sets. associated with the blasthole approach is the assaying cost.
The block outlines at the cut-off grade for all the possible data sets However, the angle RC approach must be able to offset the cost
were then used as the defined ore polygons (Figure 6). of drilling the holes since they will only be drilled if needed for
ore control.
As shown in Figure 8, the blasthole generated polygons
consistently outperformed the angle RC polygons. Further
testing was completed to determine the optimal spacing for the
angle RC holes if that approach was to be applied. For the angle
RC data set, tests were also run at 9 m × 9 m spacing, 5 m × 10 m
spacing, 6 m × 6 m spacing and 4 m × 4 m spacing. Figure 8
presents results for all the angle RC tests, as well as estimated
results (linearly interpolated) for other configurations as
compared to the 6 m × 6 m blastholes. As shown in the graph, the
gross revenue increases as the spacing is reduced, reflecting the
benefit of the added information (drill holes). For the angle RC,
offsetting the benefit is the cost of the drilling, resulting in an
optimal spacing of 5 m × 10 m. Both the 6 m × 6 m and
7 m × 7 m spacings appear to give similar results on a net
revenue basis. Due to the cost savings from using blastholes, the
net revenue from the blasthole polygons exceeds all outcomes
FIG 6 - Kriged SMU blocks (A) and ore polygons created at 0.6 g/t from the angle RC data sets. On a gross revenue, the 6 m × 6 m
outline (B) from one possible data set. blasthole data is approximately equivalent to the 5 m × 10 m and
7 m × 7 m angle RC data sets.
Revenue from SIM model
Manual ore control test
19.0
The manual approach is far more time consuming to complete

than the automated polygons, thus it was completed on only one 18.5
each from the 36 blasthole data sets and 25 angle RC data sets.
Each data set used was chosen randomly. The data sets were used 18.0
to interpolate a 2 m × 2 m block model using ordinary kriging,

Revenue x 1,000,000
then polygons were manually to separate ore and waste at 0.6 g/t 17.5
(Figure 7).
17.0
16.5
16.0
Angle(6)
blast(25)-20%
15.5 Angle(25)
Angled(25) 9m
15.0
1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35
Model
FIG 8 - Tabulation of results from blasthole and angle RC data sets.
Manual results
As with the automated approach, the polygons defined on the
blasthole and angle RC data sets were used to tabulate simulated
nodes within the respective polygons and compare results. The
FIG 7 - Manually drawn polygons. (A) Blastholes and polygons RC polygons predicted a higher revenue than that predicted by
and (B) angle RC and polygons. the blasthole polygons by about three per cent. However, the
angle RC approach resulted in polygons with more tonnes at
lower grade as compared to the blasthole data. In addition, the
ORE CONTROL RESULTS angle RC over-predicted grade as compared to the underlying
The results from the two approaches reached the same overall simulated nodes by five per cent, whereas the blasthole predicted
conclusion that the use of blasthole-based ore control would grade was overestimated by two per cent. This resulted in the
perform better, as well as insights into questions frequently blasthole approach having a net revenue of approximately
confronted in setting up ore control procedures. five per cent more.
I DOUGLAS et al
TABLE 1
Initial reconciliation results.
F1 – OC/exploration F2 – Mill/OC F3 – Mill/exploration

Tonnes Grade (g/t) Ounces Tonnes Grade (g/t) Ounces Tonnes Grade (g/t) Ounces
Oct 2006 - Mar 2007 101% 90% 91% 102% 102% 104% 102% 92% 95%
Apr 2007 - Jun 2007 96% 91% 87% 97% 109% 106% 93% 99% 93%
Initial ore control approach the mill has continued to report higher grade than defined by the
ore control system (approximately six per cent higher in the
Due to the uncertainty of sampling blastholes in wet zones and 12 months to August 2008), the overall reconciliation to the
properly sampling coarse gold especially in blastholes, the initial exploration model has been showing the model continues to
decision was to adopt the angle RC approach after completion of under-predict tonnages and over-predict grades, but resulting in
mining in saprolite. After sufficient production information was approximately five per cent higher contained metal over the
obtained, the ore control approach was revised. 12 month period.
PRODUCTION RECONCILIATION RESULTS OTHER RESULTS

As predicted by the simulation results, the initial results from the The automated ore control tests also investigated the impact of the
applied ore control system resulted in more tonnes at lower grade number of samples used in ore control. When originally
than expected. Further studies were required to attempt to pinpoint comparing the simulation re-blocked to SMU size blocks, the
whether the poor reconciliation was due to poor performance of kriging plan suggested that to achieve the correct grade tonnage
the long term model, due to poor ore control application, or both. distribution, a restricted number of samples (six) should be applied
Initially, the ore control system was completed on blocks in the ore control, or risk over-smoothing the results. Tests were
approximately the same size as the SMU (5 m × 5 m blocks), run for both the blasthole and the RC data, comparing results
which was thought to result in high dilution as well as poor between using six, 15 and 25 composites. When comparing the
routing. Improvements to the ore control system included using resulting outcomes of the simulated nodes within the polygons, the
smaller blocks in the underlying ore control model, and the use of 25 composites was superior by approximately 0.5 per cent
application of probability shells to reduce smoothing. However, on revenue as compared to 15 composites and up to three per cent
the grade tonnage distribution did not improve markedly, and the as compared to using six composites. The use of grade envelopes
conclusion drawn was that it was unlikely that 12.5 m spaced RC on the angle RC data set was also investigated in the manual test.
data could reliably predict ore and waste at the 3 m selectivity Grade envelopes were drawn in section and plan on the angle RC
being mined. After approximately nine months of mining data set prior to kriging in ore control at 0.6 g/t and 0.4 g/t.
production, the blasthole versus angle RC study was reviewed, and Polygons were then defined on the resulting kriged estimates.
a program was proposed to test whether or not blasthole sampling While the 0.6 envelope approach did define higher grade material,
would improve reconciliation. Blasthole (BH) sampling was tested the envelope caused the grade estimate to be severely overstated
and a methodology developed for sampling. Comparisons to when compared to the simulated nodes by approximately 20 per
existing RC data showed that BH data was defining ore and waste cent, and to route excessive ore tonnage to waste, resulting in
zones missed by RC drilling, and that the assays were unbiased significantly less revenue. When compared to completing ore
when paired with closest angle RC assays. An underlying control without predefined envelopes, the results from no envelope
assumption is that blasthole samples are unlikely to indicate ore were superior to the 0.4 g/t envelope approach by two per cent on
where no ore exists and vice versa, even if the sampling of the revenue, and seven per cent better than using the 0.6 g/t envelope.
blastholes was less than optimal. In both cases, the reason for the superior results was routing less
Sufficient blasthole assays were acquired by January 2007, and ore to waste.
a decision was made to move to full blasthole sampling. The ore
control system was changed to use blasthole assays in March
2007. Within three months, a measurable change in reconciliation CONCLUSIONS
was documented, as shown in Table 1. Reconciliation was tracked The use of conditional simulation was successful in predicting the
in three ways: outcomes from alternate ore control approaches. This has been
1. F1 comparing ore control to exploration model, evidenced by the following. Firstly, field tests and Mine to Mill
reconciliation results. Secondly, blasthole-based ore control has
2. F2 comparing mill to mine, and delivered to mill fewer tonnes at a higher grade – as validated
3. F3 comparing mill to exploration. through reconciliation. Thirdly, errors in blasthole assaying are
being minimised through appropriate kriging plans in ore control
While tonnage defined in ore control did decrease by though the usage of sufficient samples to overcome the inherent
approximately five per cent as predicted when comparing the randomness in blasthole sampling; and lastly, measurable cost
exploration model to ore control, there was not an associated savings have occurred by eliminating the angle RC drilling cost
increase in grade, with the grade increasing only one per cent, and reducing the tonnage to be processed. While in this study the
resulting in an apparent decrease in contained ounces by blasthole approach is favoured over the angle RC approach, it is
approximately four per cent (F1, Table 1). However, when the acknowledged that results could be very different if ore waste
reconciliation of the mine to the mill is viewed, there was a seven boundaries were more continuous, predictable and visually
per cent increase in grade delivered to the mill as compared to that mappable, as is seen in many shear zone type gold deposits.
predicted by ore control (F2, Table 1). Thus the expected increase Newmont will continue to review its ore control approach by
in grade not realised in ore control was in fact seen in the mill, closely monitoring Mine to Mill reconciliation to improve results.
leading to the conclusion that the ore control was underestimating
grade. When combined, and comparing the mill back to the
exploration model, (F3, Table 1), there is a clear decrease in REFERENCES
tonnage and an increase in grade. While three months is not a Deutsch, C V and Journel, A G, 1998. GSLIB, Geostatistical Software
sufficient time frame to have high confidence, it was an indication Library and User’s Guide, second edition, 369 p (Oxford University
that the ore control change was having the desired effect. While Press: New York).
HOME
Impact of Bench Height, Cut-Off Grade and Production Spacing

on Misclassification in Ore Control – Challenges Produced by
High Nugget, Low Continuity Ores
W Hardtke1, L Allen2, B Perry3 and I Douglas4
ABSTRACT
Model reconciliation problems at the low-grade, high nugget Phoenix
gold-copper deposit led to a thorough examination of the influence of
production decisions on ore waste classification. A conditional
simulation study of the deposit was used as the basis for: quantification
of dilution at different bench heights, quantification of ore waste
misclassification relative to bench heights and cut-off grades, and
determination of an appropriate number of samples to optimise ore
control polygon definition.
Actual reconciliation data were empirically used to support
conclusions. This paper presents the results of the study and analyses the
relationship between misclassification of ore waste blocks and cut-off
grade. The paper also explores the implications on long-term mine
planning.
INTRODUCTION
The Phoenix deposit is a gold-copper deposit located in Northern
Nevada, USA. Gold mineralisation is characterised as low-grade, FIG 1 - Location map (Newmont Nevada Properties).
but with a significant coarse gold component controlled by
localised high-angle structures within broad, shallow-dipping
sedimentary formational controls. The inherent variability of this occurs within the five kilometres long, 60 to 70 W-dipping,
deposit has resulted in particular challenges in regards to north-south Copper Canyon structural corridor with cumulative
reconciliation between the resource model, ore control and the apparent normal displacement of 650 to 825 m as described by
mill. This is reflected by high nugget effects, low continuity in Newmont in the annual Phoenix reserve report (2006). This 250
the variography and a tendency to overestimate the gold grade in to 550 m wide corridor is bound on the west by the Canyon
the resource model. fault, with 600 m of apparent normal offset, and to the east by
the ore-controlling Virgin fault, with 100 to 250 m of apparent
Operations have recently changed from 6 m to 12 m bench
normal offset. Stratiform skarn mineralisation is cut by
height to improve rock fragmentation, which has resulted in
high-angle west-dipping faults. Au, Cu and Ag reserves are
improved mill throughput. However, with the 12 m bench height,
contained mostly in the Pennsylvanian-Permian Antler overlap
attempts to raise the head grade by increasing the cut-off did not
sequence, which generally consists of a fining upward sequence
result in a higher grade at the mill. Conditional simulation was
of ferruginous conglomerate of the Battle Formation, limestone
used as a tool to quantify the difference in dilution and
of the Antler Peak Formation and calcareous sandstone of the
misclassification caused by the larger bench height. It was also
Edna Mountain Formation. The Cambrian Harmony Formation,
used to optimise ore control parameters such as the number of
consisting of sandstone and siltstone, is unconformably overlain
blasthole samples and the use of cap grades.
by the Antler sequence rocks and is only weakly mineralised.
Allochthonous argillite and chert of the Permian Havallah
LOCATION AND GEOLOGICAL SETTING sequence overlie the Antler sequence rocks in the upper plate of
The Phoenix deposit is located near Battle Mountain, Nevada, the Golconda thrust.
USA (Figure 1). The deposit area was previously mined for Gold-copper reserves at Copper Canyon are roughly divided
copper and gold by Duval Corporation between 1964 and 1982 into two by the Copper Canyon stock. Pyrrhotite-pyrite skarn of
and by Battle Mountain Gold for gold beginning in 1984. The the Antler Peak Limestone hosts high-grade gold ore in the
deposit came under Newmont control through the 2002 merger Fortitude orebody north of the stock, whereas lower gold grades
between Newmont and Battle Mountain Gold. Mineralisation are present in the underlying Battle Formation. In contrast,
Au-Cu-Ag mineralisation in the Bonanza area south of the
stock is hosted mainly in skarn of the Battle Formation and
1. Senior Consulting Geostatistician, Newmont Mining Corporation,
10101 E Dry Creek Road, Englewood CO 80112, USA.
particularly in the lowermost conglomerate unit. Only a small
Email: [email protected] amount of Antler Peak Limestone is recognised south of the
stock and no Edna Mountain Formation is present.
2. MAusIMM, Director of Reserves, Newmont Mining Corporation, Fracture-controlled Au-Cu mineralisation occurs along the
north, south and west margins of the Copper Canyon stock in
hornfelsed Havallah sequence in the West Bonanza, Reona and
3. Senior Consulting Geostatistician, Newmont Mining Corporation, Sunshine deposits, respectively.
CONDITIONAL SIMULATION FRAMEWORK
4. Group Executive – Reserves and Geostatistics, Newmont Mining
Corporation, 10101 E Dry Creek Road, Englewood CO 80112, USA. Conditional simulation studies have traditionally been used to
Email: [email protected] evaluate uncertainty as discussed by Goovaerts (1997). However,
W HARDTKE et al
the use of a single simulation is useful to provide a ‘ground-truth’

representation of a deposit. In order to investigate ore-waste
classification and misclassification issues, a conditional
simulation was created to represent the F1 area at Phoenix. This
area was selected because of the quantity of production data
available and because it had relatively high gold and low copper
grades. Blasthole data from 6 m benches were used to condition
the simulation, with only six and 12 m bench heights tested. The
4.8 m spaced blasthole data provided the necessary resolution
essential for examining misclassification at a detailed scale.
Several conditional simulations were developed for both gold
and copper within the F1 deposit using the sequential Gaussian
method in the GSLIB program SGSIM as presented by Deutsch
and Journel (1998). The variogram models and the search criteria
were developed using the Isaaks & Co Sage program (2001).
Nodes were simulated on a 1.2 m × 1.2 m horizontal spacing and
a 6.0 m vertical spacing. A single simulation was then selected to
represent the ground truth for the F1 deposit. This simulation
was selected because it not only had the same mean and variance
as the blasthole data but was visually most representative. The
visual validation was very important because the high nugget
effect tended to minimise the continuity of the mineralisation. A
typical bench from the simulation is shown in Figure 2.
ANALYSIS OF PRODUCTION DECISIONS USING

SIMULATED NODES
The Phoenix deposit has some difficult mining issues because it
is low-grade with coarse gold in ore zones that have only limited
continuity at the cut-off grades required for the flotation mill.
The combination of the inherent variability of the deposit,
sampling problems due to coarse gold and a relatively high
cut-off grade have led to excessive misclassification of ore and
waste. Possible solutions for these issues would include cost
reduction measures to lower the cut-off grade or increasing
FIG 2 - Plan view of selected simulation.
efficiencies in ore control by reducing the misclassification. With
the use of conditional simulations it is possible to quantify the
dilution associated with larger bench heights and as well as the cut-off grade and affects block revenues, copper is generally
misclassification due to the intrinsic problems caused by the ignored in this paper as the degree of ore waste misclassification
variability of the deposit. results almost entirely from gold variability.
In order to reduce operating costs, the bench height at Phoenix
was recently increased from six to 12 m. This change resulted in Quantification of dilution at different
better rock breakage and an increase in mill throughput but it bench heights
was not clear whether the reduced operating costs were sufficient
to offset the lost revenue due to dilution. Because of the The simulated nodes were re-blocked into Selective Mining Unit
difficulty in determining the grade of a given ore control block, (SMU) size blocks for both six and 12 m bench heights. An SMU
the conditional simulation was used to determine the relationship of 9.75 m × 9.75 m was selected for the 6 m bench height and
between bench height and misclassification. Not only can 19.5 m × 19.5 m for the 12 m bench height. A summary of the
misclassification be affected by bench height but it is sensitive to resulting tonnes, grade and net revenue is shown in Table 1. At a
cut-off grade and ore control parameters such as the number of 0.51 g/t cut-off, which is the approximate mill cut-off with no
samples used in kriging. Copper mineralisation is fairly uniform copper credit, the revenue for the 6 m bench is 62 per cent higher
in the F1 area and can be characterised as low-grade and and appears to be a substantially better choice. This significant
diffusive. While important in the analysis as it influences the difference in revenue is entirely based on geometric dilution but
TABLE 1
Comparison of 6 m and 12 m bench heights.
Cut-off 6 m bench height 12 m bench height

g/t Tonnes Grade g/t Net revenue Tonnes Grade g/t Net revenue
0.000 30 492 612 0.336 (90 095 443) 30 492 612 0.336 (90 095 443)
0.171 22 789 890 0.435 (28 282 484) 23 996 588 0.422 (35 880 458)
0.343 13 375 257 0.559 10 864 270 15 046 755 0.507 (334 741)
0.514 5 966 714 0.727 21 931 343 5 968 716 0.645 13 537 228
0.686 2 572 596 0.905 17 253 652 1 466 549 0.823 7 718 050
0.857 1 155 912 1.087 11 246 768 410 995 0.984 3 290 732
1.029 522 622 1.275 6 768 782 89 363 1.179 1 009 303
IMPACT OF BENCH HEIGHT, CUT-OFF GRADE AND PRODUCTION SPACING ON MISCLASSIFICATION IN ORE CONTROL
it does not take into account misclassification due to the selective mining units) and the kriged ore control blocks is caused
challenge of correctly predicting the block grade. by misclassification of ore and waste due to imperfect knowledge
The larger the SMU size the easier it is to predict the grade of from the spacing of the blasthole data, or information effect. The
that block, and ultimately classify it as either ore or waste, which poor but uniform results of the 16 models indicate that the actual
should increase the revenue. However, at the same time the location of the blasthole data is unimportant statistically. This
implies that the continuity of the gold mineralisation is nearly
grades of the blocks are also decreasing due to dilution, which
random at the current 4.8 m blasthole spacing.
lowers the revenue of the deposit. It appears that within the
normal range of cut-offs at Phoenix the dilution caused by larger One of the 16 blasthole models is plotted against the ‘ground
benches is much more costly than the revenue generating affects truth’ (re-blocked simulation) in Figure 3, showing what the
of better ore waste classification. misclassification looks like. The blocks are divided into four
quadrants:
Quantification of ore waste misclassification 1. the lower left is waste that is modelled as waste,
As there is only one geologic domain (Upper Battle Fm) within 2. the lower right is waste that is modelled as ore,
the F1 area the determination of the ore waste boundary and 3. the top right is ore that is modelled as ore, and
subsequent classification of blocks into either ore or waste was
based entirely on the grade estimation of the SMU blocks. A 4. the top left is ore that is modelled as waste.
methodology was developed to determine the variability This graph shows the difference between perfect selectivity,
associated with Kriged ore control estimates: which would be the blocks above the horizontal line
• Begin with a simulation node spacing that is one-fourth or (4349 blocks at 0.72 g/t), the predicted grade, which is the
less of the ore control blasthole spacing (a node spacing of modelled blocks to the right of the vertical line (4298 blocks at
1.2 m × 1.2 m was used). 0.73) and the actual grade (4298 blocks at 0.66 g/t) which is the
true grade of the blocks to the right of the vertical line.
• Extract simulated 4.8 m × 4.8 m blasthole patterns. In this
study 16 patterns were created using 1.2 m offsets.
Appropriate number of samples for ore
• Krige the SMU size block using the extracted simulated
blastholes. This was done for the 9.75 m × 9.75 m × 6 m and control definition
the 19.5 m × 19.5 × 12 m SMU blocks. The importance of misclassification in ore control is made very
• Compute the tonnage and grade for each of the 16 different clear in Table 2, which indicates that approximately half of the
models. possible net revenue from the F1 area is being lost. In order to
improve material classification the first step is to understand
• Compare the models with the ‘ground truth’ simulation. what is causing the problem. Unfortunately, the root of the
Table 2 shows the comparison of the 16 models (6 m benches) problem is the inherent variability of the deposit, which is
with the simulation at the 0.5 g/t cut-off. Each of the ore control believed to be too great to accurately predict block grades using
model estimates yields a noticeably lower tonnage and grade, and 4.9 m blasthole data. With the current blasthole spacing there are
hence net revenue, than the original re-blocked simulation. This four blastholes per SMU block (6 m benches) and the optimal
difference between the re-blocked simulation (perfectly known number was investigated. Figure 4 shows the probability of
TABLE 2
Blasthole kriging at 4.9 m spacing versus re-blocked simulation.
Model Tonnes Grade Au (kg) Revenue Process Net revenvue
Sim 6 280 668 0.7851 4941 83 401 500 54 001 342 29 400 158
Model Tonnes Grade Au (kg) Revenue Process Net revenvue
1 5 990 306 0.6612 3965 66 918 075 51 504 804 15 413 271
2 5 843 979 0.6646 3878 65 453 325 50 246 680 15 206 645
3 5 829 681 0.6646 3874 65 395 050 50 123 744 15 271 306
4 5 684 174 0.6646 3786 63 903 000 48 872 678 15 030 322
5 5 823 085 0.6612 3857 65 100 000 50 067 038 15 032 962
6 5 813 983 0.6680 3883 65 535 750 49 988 773 15 546 977
7 5 726 926 0.6680 3832 64 684 725 49 240 253 15 444 472
8 5 765 210 0.6680 3845 64 904 175 49 569 421 15 334 754
9 5 952 819 0.6612 3935 66 417 750 51 182 492 15 235 258
10 5 760 609 0.6680 3849 64 971 900 49 529 867 15 442 033
11 6 044 293 0.6578 3986 67 278 225 51 968 990 15 309 235
12 5 910 431 0.6612 3907 65 940 000 50 818 037 15 121 963
13 5 690 320 0.6680 3803 64 197 000 48 925 516 15 271 484
14 5 871 590 0.6646 3903 65 886 450 50 484 080 15 402 370
15 5 876 538 0.6612 3894 65 726 325 50 526 622 15 199 703
16 5 730 706 0.6646 3821 64 486 800 49 272 756 15 214 044
Mean = 5 832 166 0.6642 3876 65 424 909 50 145 109 15 279 800
Stdev = 105 756 0.00328 55
W HARDTKE et al
FIG 3 - Misclassification of blocks at a 0.51 cut-off and 6 m bench height.
from an increased number of samples. These blocks are correctly

classified about 55 per cent of the time using 16 samples which
increases to close to 80 per cent when using 50 samples.
In ore control, not having a sufficient number of blasthole
assays within a SMU block to reliably predict its grade near the
cut-off is augmented by adding up to 40 of the surrounding
blastholes. This is not as desirable as having more internal data
points but still increases the probability of classifying blocks
correctly. It also tends to decrease the mean grade of those blocks
which subsequently leads to lower predicted polygon grades as
increasing the number of blasthole assays smooths the estimate.
Comparing the effects of misclassification in 6 m

and 12 m SMU blocks
In order to decrease mining costs while increasing ore production
and mill thoughput, the bench height at Phoenix was increased
6 m to 12 m. As expected, the head grade dropped due to the
added dilution of the larger SMU size. Because the lower head
FIG 4 - Probability of misclassifying blocks with different number of grade was incompatible with the yearly production goals the
data points. cut-off grade was increased to offset the affects of the added
dilution. However, the higher cut-off grade did not result in a
higher head grade. There can be a significant difference between
classifying three randomly selected blocks (6 m) correctly at a the head grade assuming perfect selectivity and the actual grade
0.51 g/t cut-off with a varying number of data points (simulated which is a combination of ore and waste that has been
nodes). Each of these blocks was sampled 1000 times using one misclassified as ore. Figure 5 shows how the amount of
to 50 internal data points to determine the block grade. While misclassification changes with the cut-off grade and how different
this figure shows blocks from the 6 m model it suggests that that amount is for 6 m and 12 m benches.
when comparing the 6 m bench height which has four blastholes
The difference between the 6 m actual and the 12 m actual is due
per SMU block at the current blasthole spacing with the 12 m
to dilution and there are very few 12 m blocks with grades
bench which has 16 blastholes, the 12 m bench should do a much
better job of classifying the block correctly, especially routing greater than about 0.6 g/t. Figure 5 also shows that the 12 m
the ore blocks as ore instead of waste. Figure 4 implies that the benches significantly over predict the grade relative to the 6 m
degree of misclassification can be somewhat mitigated by the use benches. As the cut-off grade increases, the over prediction
of a sufficient number of samples. At about 16 samples the continues to increase while the actual grade remains more or less
predictive abilities of the data stabilise for the high grade constant due to dilution and misclassification in the 12 m bench
(0.80 g/t) blocks with the rate of correct classification being near case. This figure shows clearly why the increase in cut-off grade
100 per cent. Low-grade (0.31 g/t) blocks are incorrectly did not result in a significantly higher grade to the mill.
classified about 23 per cent of the time with 16 samples and The predicted grade can be calibrated so it is the same as the
drops to about ten per cent by increasing the number of samples actual grade with the implementation of grade capping. Figure 6
to 50. Blocks closer to the cut-off grade see a greater benefit indicates that the predicted grade is the same as the actual grade
IMPACT OF BENCH HEIGHT, CUT-OFF GRADE AND PRODUCTION SPACING ON MISCLASSIFICATION IN ORE CONTROL
with a cap grade of approximately 3.5 g/t. This figure also shows
that the net revenue increases as the cap grade becomes higher
until it reaches about 2.5 g/t. The use of a cap grade (down to
2.5 g/t) decreases the affect of the higher grade blastholes
(smearing) and reduces the size of the ore control polygons,
resulting in an increase in revenue.
CONCLUSIONS
Misclassification of ore and waste at Phoenix can lead to a lower
than anticipated revenue regardless of the SMU size. This
misclassification can be mitigated, but not eliminated, by
increasing the number of samples used for ore control
delineation. In addition, a capping strategy should be
implemented to adjust the predicted grade to the actual grade and
to increase the net revenue.
At a cut-off grade of 0.051 g/t, the decrease in revenue because
of dilution greatly out weighs the benefit of better ore
classification when changing from 6 m to 12 m benches. While
both the 6 m and 12 m bench heights show diminishing returns
FIG 5 - Comparison of predicted, perfect and actual grades.
from an increase in cut-off grade because of misclassification,
the 12 m benches are less favourable because of extreme grade
dilution. Increases in cut-off grade will not change the mill grade
significantly.
Conversely, misclassification can be reduced by decreasing the
cut-off grade, especially when mining on the smaller bench
height. Operations should find a way to increase the
fragmentation and mill throughput for the 6 m bench height
because the cost of dilution is high. Not only is the revenue much
lower at 12 m, but 45 per cent of the potential ore is being put in
the dump at a 0.51 g/t cut-off.
REFERENCES
Deutsch, C V and Journel, A G, 1998. GSLIB, Geostatistical Software
Library and User’s Guide, second edition, 369 p (Oxford University
Press: New York).
483 p (Oxford University Press: New York).
Isaaks & Co, 2001. SAGE2001, General Purpose Software for
Variography (Users Manual), 67 p.
Phoenix Mine Geology and Mine Engineering, 2007. Mineral resource
FIG 6 - Using grade capping to adjust the predicted grade.
and ore reserve report as of December 31, 2006, Newmont Mining
Corporation internal report.
HOME
A Geology-Based Joint Conditional Simulation of a Lateritic

Nickel Deposit – Quantification of Risk in Grade Tonnage Curves
and Resource Categorisation
J A Lopes1, C F Rosas2, J B Fernandes3 and G A Vanzela4
ABSTRACT Analysis (PCA). The effectiveness of this approach is limited,

because the PCA ignores cross-correlations at distances different
The risk quantification in grade-tonnage curves is decisive for capital
investment in mining projects. Geostatistical simulations for orebodies than zero. Desbarats and Dimitrakopoulos (2000) presented an
can be used to obtain grade-tonnage curves, find uncertainty and risk extension of the PCA approach to joint simulation of multiple
assessments. Applying these in multi-element deposits can be a difficult variables, replacing PCA with the Maximum/Minimum
practice, as the related attributes using the traditional co-simulation Autocorrelation Factor (MAF). It is a factorisation method
approaches require intensive computational work and may be impractical originally developed for remote-sensing applications (Switzer
for use in the mineral industry. This paper presents the risk assessment and Green, 1984). The advantage of MAF is that it is possible to
for integrating grade-tonnage curves and resources categorisation for produce the uncorrelated factors following the linear model
lateritic nickel deposits in the central region of Brazil, by joint simulation of coregionalisation with two structures (Boucher and
of multiple correlated variables of interest – Ni, MgO and SiO2. The Dimitrakopoulos, 2007, 2009). The MAF is an approach that
joint simulation of these variables is based on Minimum/Maximum
spatially decorrelates the variables involved to non-correlated
Autocorrelation Factors (MAF). MAF is an approach based on spatial
decorrelates of the original variables involved, which are transformed to factors, which are independently simulated and back transformed
non-correlated factors. The independent factors are individually simulated to conditional simulations that reproduce the cross-correlations
and back transformed to simulate correlated deposit attributes that of the original variables. The joint simulation based on MAF
reproduce the cross-correlation of the original variables. Based on this is shown to be an effective and relativity efficient method
approach, the resources are categorised honoring Ni joint simulated of risk assessment when exiting two or more variables
results by applying the 15 per cent rule, where it is considered there will (eg Dimitrakopoulos and Fonseca, 2003; Boucher and
be a statistical error <15 per cent, with 90 per cent of a confidence Dimitrakopoulos, 2007).
interval per production period.
Jointly simulated attributes of orebodies can assist resource
categorisation, a theme widely discussed in mineral resource
INTRODUCTION evaluation. Categorising resources with conditional simulation,
Geostatistical simulation methods are used to quantify geological while considering mining production periods, is a well-known
risk, uncertainty in mineral deposits and assess risk in different approach in resources evaluation. This approach consists of
aspects of mining projects. Simulation methods for individual classifying mineral resource while considering error and
variables are frequently efficient; however, in many applications, confidence intervals using the 15 per cent rule. This implies that
reproduction of the spatial dependence and joint simulation for estimated grade, tonnage and metal content in a production
multiple-variables is critical and rarely implemented (Deutsch period will be within 15 per cent accuracy at a 90 per cent
and Journel, 1998), for example, iron, lateritic nickel and confidence level. This paper presents the risk assessment of
polimetalic deposits. Existing joint simulation methods for integrating grade-tonnage curves and resource categorisation
multivariable deposits could be applied, like conventional studies for Montes Claros de Goiás (MCG) lateritic nickel
co-simulation (eg Verly, 1993; Chiles and Delfiner, 1999) and deposits by joint simulation of multiple correlated variables of
collocated co-simulation with a Markov-type coregionalisation interest – Ni, MgO and SiO2, based on Minimum/Maximum
(Almeida and Journel, 1994), but they are too computationally Autocorrelation Factors (MAF). The mineral resource
intensive to be practical, principally when there are more then categorisation is applied on a Ni joint simulated result
two variables being considered. The computational inefficiency considering error and confidence intervals per production period.
substantially increases with the number of variables being The MCG lateritic nickel deposits are located in the west portion
co-simulated. A computationally efficient alternative was of Goiás state in Brazil (Figure 1). Given that the MCG has two
proposed by David (1988) who introduced joint simulation based different types of ore, limonitic ore and saprolitic ore, this paper
on the decorrelation of variables using Principal Component was developed only considering the saprolitic ore.
1. MAusIMM, Senior Geologist, Votorantim Metais Ltda, Dept° MONTES CLAROS DE GOIÁS DEPOSIT AND
Geologia/Exploração Mineral, Edifício Esplanada, Praça Ramos de DATA AVAILABLE
Azevedo, 206-15° andar – Conjuntos 1530/1540, Centro, São Paulo
SP 01037-010, Brazil. Email: [email protected] The Montes Claros de Goiás (MCG) lateritic nickel deposits are
2. Senior Geologist, Votorantim Metais Ltda, Dept° Geologia/
formed by a lateralisation process on ultramafic igneous rocks
Exploração Mineral, Edifício Esplanada, Praça Ramos de Azevedo, (Dunite and Piroxenite) and are defined by chemical criteria with
206-15° andar – Conjuntos 1530/1540, Centro, São Paulo SP different litotypes: – Soil – SO, Limonite – OX, Sarplite – SA,
01037-010, Brazil. Email: [email protected] Bedrock – BD and Chalcedony – CD, shown in lateritic profile in
3. Geologist, Votorantim Metais Ltda, Dept° Geologia/Exploração
Figure 2. The saprolite and limonite are mineralised types where
Mineral, Edifício Esplanada, Praça Ramos de Azevedo, 206-15° saprolitic ore is destined to the Ferronickel (FeNi) process and
andar – Conjuntos 1530/1540, Centro, São Paulo SP 01037-010, limonitic ore is destined to High Pressure Acid Leaching (HPAL)
Brazil. Email: [email protected] process. There are 26 768 one metre composites samples from
4. Senior Geologist, Votorantim Metais Ltda, Dept° Geologia/
reverse circulation drilling (RC), where 20 474 samples are
Exploração Mineral, Edifício Esplanada, Praça Ramos de Azevedo, limonitic ore and 6294 samples are saprolitic ore. Table 1 and
206-15° andar – Conjuntos 1530/1540, Centro, São Paulo SP Figure 3 show the statistics resume and histogram for Ni, MgO
01037-010, Brazil. Email: [email protected] and SiO2.
J A LOPES et al
FIG 1 - MCG deposit location.
FIG 2 - MCG lateritic profile.
A GEOLOGY-BASED JOINT CONDITIONAL SIMULATION OF A LATERITIC NICKEL DEPOSIT
FIG 3 - Histogram from Ni, MgO and SiO2.
Λ1 = Q1B Q1T (3)

TABLE 1
Statistics of one metre composites.
and Q2 is the matrix of eigenvectors from decomposition:
Ni (%) MgO (%) SiO2 (%)
Number sample 6294 6294 6294 Q2M(Δ)Q2T = Q2{½[[Γ y(Δ)]T + [(Γ y(Δ)]]}Q2T (4)
Mean 1.08 18.31 35.85
Where matrix Γ y(Δ) is an symmetric variogram matrix at lag
Median 0.96 17.95 35.15 distance Δ for regular PCA factors:
Std deviation 0.58 6.18 7.16
Variance 0.34 38.14 51.24 Y(x) = Z(x) A
Kurtosis 0.68 -0.69 2.99
where:
Asymmetry 0.89 0.16 1.16
Maximum 4.27 38.70 69.2 A = Δ-1/2
Minimum 0.06 1.28 10.8
Several Δ lag distances may be used for values lower than the
range, and the resulting eigenvectors averaged.
METHODOLOGY The joint simulation of multiple correlated variables using the
MAF approach process steps are below:
Joint simulation using maximum/minimum • normalise the available variables,
autocorrelation factors (MAF) • use MAF transformation to generate non-correlated factors
(MAF),
Multi-elements are represented by multivariate stationary and
ergodic vector random function. Considering a multivariate l • produce the variogram for each factor,
dimensional, Gaussian, stationary and ergodic spatial vector • conditionally simulate each factor,
random function:
• validate the simulation of factors,
Z(x) = [Z1(x),…, Zl (x)] T • back transform the simulated factors to normal score variables,
• de-normalise the simulated variables, and
The Maximum/Minimum Autocorrelation Factors are defined
as the l of the orthogonal linear combinations:
• validate final results.
Y(x) = aiTZ(x), i = 1,…, l Resource categorisation using the 15 per cent

rule
of the original multivariate vector Z(x). MAF are derived The mineral resource categorisation using the 15 per cent rule is
assuming Z(x) is represented by a two-structure linear model of estimated within 15 per cent of the true value, with 90 per cent
coregionalisation. The MAF transformation can be written as: confidence in the production period. This implies that the
estimated grade, tonnage and metal content in a production
Y(x) = AMAFZ(x) (1) period will be within 15 per cent accuracy at a 90 per cent
confidence level, ie nine out of ten times production periods will
and the MAF factors are derived from: have an estimation error less than 15 per cent. According to
Dohm (2004), who describes an application of this resource
categorisation criterion using threshold values, the categorisation
AMAF = Q2Λ1-1Q1 (2) criteria are as follows:
Where the eigenvectors Q1 and eigenvalues Λ1 are obtained 1. calculate the error with 90 per cent confidence:
from decomposition of the multivariate covariance matrix B of
Z(x) at lag distance: Error = 1.645 CoV / n
J A LOPES et al
where: TABLE 2
Transformation matrix.
CoV = (Std Deviation)/Mean
-0.546 0.679 0.471
n = number of blocks mined per period 0.083 0.424 -0.088
2. Change support, ie block to production quarter and block to 0.833 0.598 0.077
one year production period – calculate the variability
reduction factor:
3. FPQ = CoVBlock / CoVPQ
FPY = CoVBlock / CoVPY
where:
CoVBlock = CoV for blocks of the same size used for estimation
CoVPQ = coefficient of variation of quarter production period
CoVPY = coefficient of variation of year production period
4. Calculate the measured and indicated threshold for

classification;
Measured (Threshold) = (FPQ × 0.15)/1.645
Indicated (Threshold) = (FPY × 0.15)/1.645
5. Classifying resources:
Measured = CoVBlock ≤ Measured (Threshold)
Indicated = Measured (Threshold) < CoVBlock ≤
Indicated (Threshold)
Inferred = Indicated (Threshold) > CoVBlock
After resource classification, it is an important practice to

validate the classification with geology and draw the final
measured, indicted and inferred limits.
JOINT SIMULATION OF Ni, MgO AND SiO2
Normal-score transformation
The normal-score transformations are based on rank ordering of FIG 4 - Ni, MgO and SiO2 MAF cross-variogram.
data and decrease the influence of outliers. Transform the Ni,
MgO and SiO2 original data distribution to normalised data
distribution. This assists in the calculation of the variogram and simulation (Dimitrakopoulos and Luo, 2004). The simulation is
covariance matrix. performed on a grid of 117 620 nodes within geological limits.
Twenty realisations and validations are run for each MAF
MAF transformation and variogram analysis reproducing histogram and variogram. The validation of the
MAF simulations is not presented here. Only the final validation
The MAF transformation is calculated by multiplying a vector of
in the data space is presented.
variables Ni, MgO and SiO2 by a vector loading from the
transformation matrix AMAF (Equations 1, 2, 3), and used to
generate the three min/max autocorrelation factors for each MAF back transformations
variable (Table 2). The lag distance Δ used in transformation
Equation 4 is 15 m taken from variogram analysis on MAF The realisations of each MAF were back transformed to
values, which demonstrates decorrelation in cross variogram simulated normal score variables, using the corresponding
between calculated MAF values (Figure 4). The variograms of inverse matrix of the MAF loadings in Table 2. The normal score
saprolitic ore are performed and fitted for each MAF by for each variable, Ni, MgO and SiO2, were back transformed to
traditional spherical variogram (Figure 5). The MAF variograms data space. Three realisations and the data of Ni grade
are subsequently used for the simulation. distribution are shown in Figure 6.
Conditional simulation of MAF Validation of joint Ni-MgO-SiO2 simulations results

The conditional simulation is performed independently for each Several validation checks are performed on the results of joint
MAF, using a sequential algorithm based on sequential Gaussian Ni, MgO and SiO2 simulations using the MAF transformation.
FIG 5 - Ni, MgO and SiO2 MAF variogram.
The validation involves histograms, variograms and cross- simulation results of Ni, MgO and SiO2 in MCG saprolitic ore
variograms comparison between simulation and data. Figure 7 (Figure 9), show 20 possibilities of tonnage, Ni mean grade and
shows histograms comparison data versus simulation (one SiO2/MgO ratio considering Ni cut off. Based on these results, it
realisation randomly chosen). Figure 8 shows variogram and is possible quantify the risk of the Ni mean grade to be too low
cross variogram comparison of data versus simulation. All and/or the SiO2/MgO ratio to be too high, supporting important
validation results suggest that the reproduction of the original decisions to the project.
data by the joint simulations is good.
Based on the joint simulation results of Ni, MgO and SiO2 in
the MCG saprolitic ore, the resources were categorised
RISK ASSESSMENT AND RESOURCES considering statistical error and confidence intervals, honoring
each step described earlier and using the 15 per cent rule.
CATEGORISATION Resources categorisation using the 15 per cent rule criteria implies
that estimated grade, tonnage and metal content in a production
Lateritic nickel deposits have limitation processes caused by period will be within 15 per cent accuracy at a 90 per cent
SiO2, MgO, Fe and Al2O3 grades, depending on what process confidence level. The MCG saprolitic ore has blocks that are 25 ×
type is to be used. Nickel production by a Ferronickel (FeNi) 25 × 1 m3. These blocks were changed to support a quarter period
process is strongly influenced by the SiO2/MgO ratio. The production block size of 200 × 200 × 7 m3 and to a one year
SiO2/MgO must be up to 1.7 without copper coolers conditions production period block size of 200 × 200 × 28 m3 (Figure 10).
and up to 2.0 with copper coolers conditions. A SiO2/MgO ratio Figure 11 shows the distribution error at a 90 per cent confidence
more then 2.0 can generate operational furnace problems, and for blocks, quarter and yearly production periods. The quarter year
depending on the persistence of higher SiO2/MgO ratios, the block size was used to calculate the Measured resources and the
project can became unviable. The grade-tonnage curves one year block size was used to calculate the Indicated resources.
J A LOPES et al
FIG 6 - Ni grade distribution for three realisations.
FIG 7 - Validation – histogram comparison.
FIG 8 - Validation – variogram and cross-variogram comparison.
The MCG resources categorisation is then: with a Ferronickel (FeNi) process. Considering the FeNi process
requirement, the results of twenty realisations show acceptable
Measured = CoVBlock ≤ Measured (Threshold) SiO2/MgO ratio for any nickel cut off considered.
Indicated = Measured (Threshold) < CoVBlock REFERENCES

≤ Indicated (Threshold) Almeida, A S and Journel, A G, 1994. Joint simulation of multiple
variables with a Markov-type coregionalisation model, Mathematical
Geology, 26(5):565-588.
Inferred = Indicated (Threshold) > CoVBlock
simulation framework and application in a multivariable deposit, in
After resources were classified, some isolated measured blocks Orebody Modelling and Strategic Mining Planning, second edition
were considered indicated. Figure 12 shows a plan view with (ed: R Dimitrakopoulos), pp 345-354 (The Australasian Institute of
Measured, Indicated and Inferred resources distribution results of Mining and Metallurgy: Melbourne).
the categorisation criteria described above. Boucher, A and Dimitrakopoulos, R, 2009. Block simulation of multiple
correlated variables. Mathematical Geosciences, 41(2):215-237.
CONCLUSIONS Chiles, J P and Delfiner, P, 1999. Geostatistics Modelling Spatial
The risk analysis in grade-tonnage curves on multi-element Uncertainty, 695 p (John Wiley and Sons: New York).
deposits shows it is applicable in lateritic nickel deposits, or for David, M, 1988. Handbook of Applied Advanced Geostatistical Ore
any multi-element deposit, and should be considered in feasibility Reserve Estimation, 216 p (Elsevier: Amsterdam).
studies and in supporting important decisions concerning the Desbarats, A J and Dimitrakopoulos, R, 2000. Geostatistical simulation
project. The approach applied in this study presented an efficient of regionalised pore-size distribution using min/max autocorrelation
joint conditional simulation using minimum/maximum auto- factors, Mathematical Geology, 32(8):919-942.
correlation factors, which simplifies joint simulation in Deutsch, V C and Journel, A G, 1998. Gslib: Geostatisitcal Software
multi-element mineral deposits by decorrelating variables to Library and User's Guide, second edition, 369 p (Oxford University
independent factors that are simulated independently. The Press: New York).
comparison between simulated results versus original data shows Dimitrakopoulos, R and Fonseca, M B, 2003. Assessing risk in
excellent validation. The resources categorisation presented in grade-tonnage curves in a complex copper deposit, northern Brazil,
this study, classifying Measured, Indicated and Inferred based on an efficient joint simulation of multiple correlated variables,
Application of Computers and Operations Research in the Minerals
resources, while honoring the error and confidence interval in a Industries, pp 373-382 (Southern African Institute of Mining and
production period, is an important approach to support short and Metallurgy: Marshalltown).
long term production planning.
Dimitrakopoulos, R and Luo, X, 2004. Generalised sequential Gaussian
The MCG deposit has two types of orebodies: saprolitic, and simulation on group size and screen-effect approximations for large
limonitic ore. This study focuses only on saprolitic ore treated field simulations, Mathematical Geology, 36(5):567-591.
J A LOPES et al
FIG 9 - Grade-tonnage curves.
FIG 10 - Change nodes to three month and year production supports.
FIG 11 - Error percentage with 90 per cent confidence – plan view distribution.
FIG 12 - Resources categorisation – plan view distribution.
Dohm, C, 2004. Quantifiable mineral resource classification: A logical Verly, G W, 1993. Sequencial Gaussian Co-simulation: A Simulation
approach, in Geostatistics Banff – Book Series Quantitative Geology Method Integrating Several Types of Information in Geostatistics (ed:
and Geostatistics, 2:333-342 (Springer: New York). A Soares), 5:543-554 (Kluwer Academic Publishers: Dordretcht).
Switzer, P and Green, A A, 1984. Min/max autocorrelation factors for
multivariate special imagery, technical report 6, Stanford University,
Department of Statistics, 10 p.
HOME
Resource/Reserve Classification with Integrated Geometric and

Local Grade Variability Measures
R Dimitrakopoulos1, M Godoy2 and C L Chou3
ABSTRACT and risk underpinning resources/reserves. State-of-the-art

Categorisation of mineral resources and ore reserves as Measured,
approaches that quantify geological uncertainty and risk can be
Indicated, Inferred and Proven or Probable respectively, is a requirement used to standardise resource/reserve classification, support better
for both public reporting as well as internal company assessments of investment decision making, enable technically-defensible stock
mining assets. From the understanding of mineral resources to estimation exchange reporting and compliance with industry codes and
of ore reserves, geological uncertainty impacts heavily upon resource/ provide greater assurance to investors and protect the public.
reserve classification. Resource/reserve classification is traditionally based Traditionally, resource classification has been based on
on geometric criteria, as by the number of samples or drill holes within a geometric considerations and a block-by-block classification
volume around the deposit blocks being classified, or alternatively, some using criteria linked to the number of samples with-in a volume
form of a geostatistically-derived estimated variance. These criteria are
relatively easy to understand and implement – however, they do not or drill hole densities, or alternatively, a geostatistically-derived
account for the local variability of grades within a deposit, leading to estimation variance. These criteria are relatively simple to
misrepresentation of the confidence of estimated grades within given implement as well as being easily understood and transparent.
volumes. However, they have a main drawback as they do not account for
This paper proposes an approach founded on a stochastic simulation the local variability of grades within a deposit – this leads to
framework for quantifying geological uncertainty and risk in five steps: misrepresentation of the confidence of estimated grades within
generation of suitable realisations of the deposit at the appropriate block- given volumes and of the resulting resource/reserve
support scale, derivation of conditional cumulative distribution functions classification.
(ccdf) of block grades or attribute selected, utilisation of the conditional
coefficient of variation (CCV) of the grade ccdf within a block as the Past efforts to improve classification techniques focused on the
uncertainty measure for resource classification of each block in the development and use of alternative versions of grade estimation
orebody model, classification of resources/reserves, and evaluation and (kriging) variances, where classification patterns are based on the
production of risk profiles for estimates. This approach integrates both the size of the sampling grid (eg Saburin, 1983; Froideveau, Roscoe
traditionally-used geometric considerations, in the sense of data locations and Valiant, 1986; David, 1988; David and Dagbert, 1977; Arik,
and the in situ variability of grades. The applications in an alluvial diamond 2002 and others). As also noted above, these formulations may
deposit, epithermal gold lode and a porphyry copper deposit show that account for possible increases in precision as the sampling grid
there are natural groupings of the CCV of mining blocks that allow rational
categorisation of resources or reserves. The proposed approach is not only
size decreases – however, they do not incorporate uncertainty
transparent but also provides a technically defendable reporting of introduced by local variability in the actual data as eloquently
resource/reserve classification in a geostatistical context and may support presented by Ravenscroft (1992). Figure 1 illustrates this point,
better investment decision making. showing two identical configurations of a block and four drill
holes. On the right, the grade of the material sampled is highly
variable, whereas on the left block, the grades are much less
INTRODUCTION variable. If this were the same zone with the same variogram, the
Classification of mineral resources as Measured, Indicated and estimation variances would be the same, thus providing a false
Inferred and ore reserves as Proven or Probable is required for level of uncertainty assessment and classification of the blocks.
public disclosure as detailed in various codes such as JORC, CIM, Estimation variances do not adequately address the question of the
PERC, or SAMREC. In addition, categorisation of resources and precision of global resource estimates and even more so for global
reserves is an important contributor to a company’s assessment of resources above a cut-off (David, 1988), as would be the case if
mining assets and their grade-tonnage curves. It is widely accepted any other geometric measure was used. Simulation-based
that geological uncertainty and risk heavily impact upon the approaches have been suggested given the issues discussed above
classification of resources and reserves, starting from the (Audet and Ross, 2007; Wawruch and Betzhold, 2005; Dohm,
understanding of mineral resources and exploration programs to 2005; Duggan and Dimitrakopoulos, 2005; Murphy et al, 2004)
the ‘ore reserve’. Reliability, confidence, transparency and and need further consideration.
technical ‘defendability’ of classification are critical. Considering The shortcomings of the conventional reserve/resource
the worth of mining companies, benefits to shareholders, reporting classification may be explored in an example from a gold mine.
to financial institutions, and attracting the investment community Reserves classified as Measured, Indicated and Inferred are based
to decision-making in the order of billions of dollars, there is a primarily on the number of samples within a distance from a block
genuine need to understand and quantify geological uncertainty being geostatistically estimated and consideration of the spatial
grade continuity within identified gold lodes. Figure 2 shows part
of a lode where blocks have been classified to reserve categories
1. FAusIMM, Professor and Director, COSMO – Stochastic Mine according to the mine’s established approach. Although this
Planning Laboratory, Department of Mining and Materials
Engineering, McGill University, Montreal QC H3A 2A7, Canada.
2. MAusIMM, Golder Associates Santiago, Av 11 de Septiembre 2353,
Piso 2, Proviendencia Santiago, Chile. Email: [email protected]
3. COSMO – Stochastic Mine Planning Laboratory, Department of
Mining and Materials Engineering, McGill University, FDA
Building, 3450 University Street, Montreal QC H3A 2A7, Canada. FIG 1 - Estimation for identical sampling configurations with
Email: [email protected] different local grade variability and uncertainty.
R DIMITRAKOPOULOS, M GODOY and C L CHOU
Classification methodology
A resource classification methodology is outlined in Figure 3 and
consists of five steps:
1. generation of suitable realisations of the deposit at the
block-support scale considered appropriate,
2. derivation of conditional cumulative distribution functions
(ccdf) of block grades or attribute selected,
3. calculation of uncertainty measure(s) to be used for
resource classification based on the ccdf of each block in
the orebody model,
4. classification of resources/reserves, and
5. evaluation and production of risk profiles for estimates.
FIG 2 - Traditional reserve classification in a lode of a gold mine Measures of local spatial uncertainty
(left) and probabilities for the blocks to be ore (right).
A key part of the classification methodology is the use of
summary statistics and measures of uncertainty which are derived
approach is reasonable, relatively simple, transparent, compliant from the conditional cumulative distribution function of the grade
with regulations and technically defensible, the traditional Z(u) of a block at location u within the deposit (Figure 4). Basic
approach shows shortcomings when the classification is compared measures include the conditional variance, conditional coefficient
to probabilities of blocks being ore (see Figure 2). Large parts of of variation, inter percentile ranges and probability interval.
the lode that are classified as Indicated have over 80 per cent and The conditional variance (CV) measures the spread of the ccdf
90 per cent probabilities of being ore – thus it seems inconsistent around its mean value and is given by:
not to report them as Measured. Similarly, blocks which have been
K +1
classified as waste have, in fact, over 50 per cent probability of
being ore. Although the analysis above is confined to limited
CV ( u) = ∑[z k − zE* ( u)]2 [ F ( u; zk ) − [ F ( u; zk − 1 ) ] (1)
k= 1
aspects (for example and as per the JORC, one must also
accurately estimate the grade), it illustrates how conventional
where:
classification criteria do not include direct measures of geological
(grade) uncertainty apart from those related to the spatial zk k=1,…K are K threshold values discretising the range of
configuration of the drill holes. As a result, the conventional variation of z values
approaches may lead to different classification of Mineral zk is the mean of the class zk-1,(zk-1,zk] which in case
Resources compared to using classifications which integrate the of a within-class linear interpolation model
actual in situ spatial variability of the orebody. corresponds to zk = (zk-1 + zk)/2
K +1
In the following sections, stochastic simulation techniques are
shown to provide critical information for resource/reserve
zE* ( u) = ∑ z [ F ( u; z ) − F ( u; z
k k k − 1 )] is the discrete sum
k= 1 approximation to the expected value of the ccdf
classification complementing resource estimates, with the
understanding of the multifaceted nature of the classification The conditional coefficient of variation (CCV) corresponds to
process and the role of the competent person. The techniques the conditional standard deviation divided by the mean or relative
model uncertainty in estimated grades of orebody models and lead standard deviation. It expresses variability as a percentage of the
to an assessment of geological uncertainty as a function of both in mean, and is calculated as:
situ variability and drilling density. In addition, the simulation
framework allows for the integration of simulated geological K +1
alternatives with simulated grades and possible errors in assay ∑[z k − zE* ( u)]2 [ F ( u; zk ) − F ( u; zk − 1 )]
k= 1
data, drill hole locations and so on. The next section begins by CCV ( u) = (2)
briefly discussing limits of traditional classification approaches. zE* ( u)
Subsequently, an alternative methodology is outlined, based on
simulated conditional distributions of block- support grades and
The CCV better discriminates zones in terms of sample density
measures of their uncertainty for an orebody model. Measures of
(or availability of conditioning information). An advantage of the
uncertainty are detailed next using an example from an alluvial CCV over the CV is that it is expressed directly as a percentage of
diamond deposit. Resource/reserve classification in the same the mean and therefore filters out proportional effects.
deposit, an epithermal gold lode and a porphyry copper deposit
The inter quartile range (IQR) is defined as the difference
elucidate the related aspects and characteristics of the approach.
between the upper (q0.75) and the lower (q0.25) quartiles of the
Comments and conclusions follow. distribution:
RESOURCE/RESERVE CLASSIFICATION AND IQR( u) = q0 .75 ( u) − q0 .25 ( u) = F −1 ( u; 0.75) − F −1 ( u; 0.25) (3)

MEASURES OF LOCAL UNCERTAINTY
Conditional simulation enables a resource/reserve classification Similarly, the generalisation to any inter percentile range
methodology that simultaneously considers data density and (IPR), such as a 90 per cent confidence interval, is defined as:
orebody continuity as well as in situ variability. In this section,
key aspects of the methodology are described, including IPR( u) = q0 .95 ( u) − q0 .05 ( u) = F −1 ( u; 0.95) − F −1 ( u; 0.05) (4)
appropriate measures of uncertainty to be used for resource/
reserve classification. The methodology is illustrated using a Unlike the CV and CCV, IPR and its variations are not affected
diamond-bearing deposit in south-western Africa. by the magnitude of grades in the ccdf and are thus not sensitive
RESOURCE/RESERVE CLASSIFICATION WITH INTEGRATED GEOMETRIC AND LOCAL GRADE VARIABILITY MEASURES
FIG 3 - Diagrammatic representation of a simulation-based resource classification methodology.
hole locations are shown on the figure by triangles. The CV shows

good correlation between regions of dense sampling and low
conditional variance. However, this is not necessarily true for areas
of high thickness. The presence of a proportional effect is clearly
seen when comparing the CV and mean block thickness values,
suggesting that that the CV may not be a suitable measure of local
uncertainty. Similarly to the CV, the IQR shows good correlation
between regions of dense sampling and lower IQR values as well
as a higher uncertainty about the higher thickness values.
Generally, for highly skewed distributions the IQR may appear
more attractive than the CV as the IQR is a relative measure that
does not use the mean. A major drawback of the IQR is that it
ignores the internal distribution of probability densities, leading to
the over-representation of uncertainty. The CCV discriminates
better than the CV and IQR in terms of sample density (available
conditioning information) and, more importantly, does not exhibit
a proportional effect. As a result, it is more suitable for resource
FIG 4 - A conditional cumulative distribution function (ccdf) classification.
showing summary statistics and uncertainty measures for the
thickness of a 100 × 100 m2 block in a diamond deposit. The ccdf
is generated by calculating the associated probabilities for a given Resource/reserve classification measures
block from multiple simulated realisations.
The uncertainty measures in the previous section can be used to
define resource/reserve classification criteria. These criteria
to outliers. The IPR may also be scaled by the mean or median of require the use of threshold values, similarly to past efforts, where
the corresponding ccdf. Related measures of uncertainty are the thresholds were applied to estimation variances (eg review in
probability that a block grade, Z(u) is above, below or within a Sinclair and Blackwell, 2000). Threshold values may reflect the
probability interval. Figure 4 shows related summary statistics and error tolerance that is acceptable locally for the block estimates as
uncertainty measures for the ccdf of thickness of a 100 × 100 m2 well as globally for the resource estimate. The block estimate
block. The ccdf itself provides the model of uncertainty for the herein corresponds to the mean value (or e-type estimate). It is
value of the thickness attribute for a given block at location u possible to categorise resources with block precision and using
within the deposit. Each discrete point in the ccdf corresponds to resource models derived from multiple simulation realisations to
a simulated value. The ccdf is a cumulative histogram of all obtain the uncertainty in global resources or resources within
simulated values assigned to the simulation realisations of the production increments of interest in each category.
deposit. A continuous function is interpolated between the A first confidence interval based on the CCV involves fixed
discrete points to enable the assessment of probabilities for any threshold CCVs that are compared to the CCV. This is equivalent
ccdf value. to comparing the confidence interval about the mean of ± the
Figure 5 shows the CV, conditional mean, CCV and IQR for the standard deviation against a threshold confidence interval
same size of thickness blocks over a part of a diamond deposit. defined as a proportion of the mean. The CCV classification
The values are generated from 50 simulation realisations and drill criterion are given by:
FIG 5 - Plan view from left to right, the conditional variance (CV); conditional mean (CM); conditional coefficient of variation (CCV); and inter
quartile range (IQR) for 100 × 100 m2 block thickness in a diamond deposit. Note that denser drilling does not ensure less uncertainty.
⎧ Measured if CCV < Threshold 2 CASE STUDIES

⎪
CCVClassification = ⎨ Threshold 2 ≤ Indicated if CCV < Threshold 1(5)
⎪ Inferred if CCV ≥ Threshold 1 A simple example in a diamond deposit
⎩
As an example, consider a CCV classification with a threshold For an alluvial diamond deposit where the important classification
CCV of 0.5. In this case, all blocks for which the confidence parameter is thickness, Figure 6 shows a plan view of the resource
classification as Indicated or Inferred using three classification
interval bounded by ± the standard deviation is over 50 per cent
criteria (CCV with a CV threshold of 40 per cent, RPR with a
of the value of the mean will be classified as inferred. The 90 per cent confidence interval and a single (for simplicity)
remaining blocks are classified as Indicated or Measured. The threshold of 40 per cent, and CCV with a CV threshold of 50 per
selection of threshold values is deferred until a later section. cent). Table 1 details the potential classified resource in this
A second confidence interval based on 90 per cent probability diamond deposit and related estimates with their uncertainty.
(a relative percentile range or RPR) is a confidence interval about Using the CCV classification, both the volume classified as
the mean that is compared with a threshold confidence interval Indicated and the associated uncertainty, as measured by the
defined as a proportion of this mean. The CCV classification is differences in maximum and minimum volumes, increase with
defined based on the standard deviation, here the confidence threshold, as expected. The increase in volume is considerable
interval is defined considering the relative position of the mean while the increase in uncertainty in relation to the expected value
to the fifth and 95th percentiles, termed negative and positive is relatively small. In global terms, the uncertainty in the volume
differences. The symmetric confidence interval used in this classified as Indicated is relatively low. The results indicate that in
criterion is: this study the 40 per cent threshold may be considered conservative.
For inferred resources, as the threshold increases the volume
⎡ z*E ( u) − F −1 ( u;0. 05) z*E ( u) − F −1 ( u;0. 95) + ⎤
decreases and associated relative uncertainty increases. The
Confidence interval = ± ⎢ *
w− + w ⎥(6) decrease in volume results from blocks moving to the Indicated
⎢ zE ( u) z*E ( u) ⎥ category. The increase in uncertainty is expected as it reflects the
⎣ ⎦
higher uncertainty related to higher threshold values. The RPR
classification when compared to the CCV classification approach,
where: classifies considerably fewer blocks as Indicated at all thresholds
w- and w+ are weights for the relative differences to the fifth and and does not represent a good assessment of variability. As with
95th percentiles, respectively. The weights facilitate asymmetric the IQR measure discussed earlier, the RPR classification ignores
confidence intervals and are determined based on the probability the internal distribution of probability densities and leads to the
intervals: over-representation of risk. It is therefore very conservative in
classifying blocks. The results presented in Table 1 show that the
[ 0.05 − F ( u; zE* ( u))] [ 0.95 − F ( u; zE* ( u))] uncertainty analysis on the resource based on three different
w− = w+ = (7) threshold values, may provide an indication as to which threshold
0.9 0.9 value could be more appropriate for the deposit. Understanding
the meaning of the threshold value is critical in applying the
Note that setting w- and w+ to 0.5 corresponds to defining the proposed classification criteria. Mining and processing parameters
confidence interval as ± the average between the negative and may further be considered in addition to external considerations
positive differences. The classification criterion is: that may be taken on board, relevant to project economics or a
company’s risk aversion level. Note that the analysis herein
⎧ Measured if Threshold < Confidence interval 2 corresponds to propagating local block uncertainty to the
⎪
RPR Classification = ⎨ Confidence interval 2 ≤ Indicated if threshold < Confidence interval 1 uncertainty in the global resource. Selection criteria of threshold
⎪ Inferred if threshold ≥ Confidence interval 1 values based on deposit characteristics are discussed next.
⎩
TABLE 1
Resource classification of 100 100 m2 thickness blocks based on CCV and RPR classification criteria using 40 per cent, 50 per cent
and 60 per cent CV thresholds.
Threshold CCV classification RPR classification

40% 50% 60% 40% 50% 60%
Category Indic Infer Indic Infer Indic Infer Indic Infer Indic Infer Indic Infer
Average volume (M m3) 3.15 3.91 5.26 1.80 6.67 0.46 0.45 6.61 2.06 5.00 3.94 3.13
Max volume (M m3) 3.02 4.27 5.66 2.00 7.17 0.55 0.50 7.16 2.19 5.45 4.21 3.43
3)
Min volume (M m 3.38 3.60 4.96 1.64 6.25 0.36 0.40 6.25 1.91 4.70 3.74 2.89
against the CCV for 25 block support realisations of the orebody.

The figure shows a general and distinctly increasing spread (risk)
in ore tonnage and average grade as the CCV increases, which is
particularly strong in the ore tonnages. This trend is also seen in
Table 2 detailing CCV and summary statistics for metal content in
the gold lode. The characteristics of the spread and slope of the
curves in Figure 7 and values reported in Table 2 can be used to
guide selection of thresholds for classification. For example,
Figure 7 suggests that there is little risk for a CCV up to around
0.65 as the related tonnage above the cut-off curves are mostly
overlapping and the average grade above cut-off appears to have
stabilised. In addition, the figure shows that after a CCV of about
1.1, the spread of tonnages and grades is at its greatest and
remains stable for higher CCV values. Hence, the higher CCVs
can be selected for the classification of the high-risk resource.
In an alternative analysis, Table 2 shows a CV with a
minimum of 4.2 per cent at the 0.65 CCV threshold, suggesting
that this threshold corresponds to the most certain measure of
metal quantity and would be a reasonable threshold to
FIG 6 - Resource classification using the CCV classification
differentiate between Measured and Indicated reserves. The
criterion with a CV threshold of 40 per cent (left); RPR
differentiation between Indicated and Inferred reserves may be
classification within 90 per cent confidence interval and a CV
based on a CV of about 6.7 per cent or CCV of 1.1, which
threshold of 40 per cent (middle); and CCV classification using a
CV threshold of 50 per cent (right).
identifies the highest end of risk. At a higher threshold only a
small amount of additional metal would be delivered. For a CCV
value between 0.65 and 1.1, both metal quantity and uncertainty
increase substantially. Hence, there is a trade-off between gains
Selecting thresholds at a gold mine in metal quantity and increase in risk when differentiating
between Indicated and Inferred reserves. Note that Table 2 allows
The example of an open pit gold mine in Western Australia is used straightforward reporting of average metal for each category as
to further explore CCV based classification. Issues addressed well as a confidence interval, minimum, maximum and
include the selection of threshold values for reserve classification. proportion of range, as in Table 3 (the ± in the table are in terms
The example is comprised of a set of epithermal gold lodes in a of ranges between maximum and minimum simulations). Also
zone of the deposit which are to be classified as Measured, note that the tolerance used in this case study could be lower
Indicated or Inferred resources. Figure 7 shows tonnage of ore resulting in possibly expanding the Measured resources. Table 3
above cut-off (1 g/t) and average grade above cut-off plotted provides the detailed reporting of the reserve classification. An
FIG 7 - Average grade above the 1 g/t cut-off (left) and tonnage of ore above cut-off (right) plotted versus the CCV for 25 realisations
of a gold load.
TABLE 2 spread (risk) in ore tonnage as the CCV increases. Figure 8

suggests the presence of little risk for a CCV up to about 0.45 as
CCV and summary statistics for metal content in a gold load.
the related tonnage above the cut-off curves are mostly
CCV Min Max Range Average CV Increase overlapping. The spread of the same curve increases up to a CCV
metal metal metal (%) in metal of 0.85, when the spread appears to reach its maximum and
(oz) (oz) (oz) quantity stabilises at that maximum spread. These two values of the CCV
0.35 576 1151 575 812 17.0 can then be used as the threshold values chosen for resource
classification.
0.40 2925 4645 1720 3849 10.3 3037
0.45 6721 9834 3113 8840 7.8 4991 Comments
0.50 14 245 20 000 5755 18 120 6.3 9280
It is important to note that the thresholds used in the above
0.55 29 146 36 105 6959 33 353 4.4 15 233 examples will change from deposit to deposit, different geological
0.60 48 050 58 724 10 674 53 916 4.3 20 563 zones and different commodities. However, the process of
generating CCV values will remain the same and the relatively
0.65 72 354 83 978 11 624 77 909 4.2 23 993
distinct brakes in the various curves and corresponding tables will
0.70 97 457 116 560 19 102 107 747 4.7 29 838 be present. In addition, it is important to reiterate that unlike
0.75 127 570 155 776 28 206 140 369 5.3 32 622 traditional approaches, classification of the blocks using any of the
proposed measures generated through use of simulations reflect
0.80 151 644 187 932 36 288 168 126 5.4 27 757
both data density and local grade variation. Figure 9 shows the
0.90 191 436 238 876 47 440 212 314 6.0 44 188 CCV classification of blocks in a section of the gold lode
1.00 212 766 270 189 57 423 238 691 6.4 26 377 discussed above, along with the available drilling data. Although
drilling density varies, the classification does not always show a
1.10 223 778 286 719 62 941 251 968 6.7 13 277 strong correlation with drilling density. For example, a densely
1.20 228 104 295 563 67 459 258 527 6.9 6559 drilled area may not always be classified as Measured, as
1.30 229 903 298 081 68 178 261 089 6.9 2562 traditional classification approaches may imply.
1.40 230 305 299 453 69 148 262 317 7.0 1227
Blocks reported in resource classification should be of the
same support size. If this is not the case, support-independent
1.50 230 549 299 779 69 230 262 641 7.0 324 measures may be derived. A known approach is to standardise
the CCV of a block of a given support size using the square root
of the corresponding dispersion variance (eg Sabourin, 1983).
Another practical comment is that the number of simulations
TABLE 3 required in this approach depends on the sensitivity of the results.
Reporting of classification and evaluation of ore reserves for If additional simulations do not effect the overall classification,
metal, tonnes and average grade. they are not required. All classification approaches may be
Category CCV Tonnage Average Metal
improved by choosing a rational spatial grouping of classified
(t) grade (g/t) (oz) blocks in a category to avoid their unrealistic dispersion. Last but
not least, the ability to efficiently simulate large deposits and
Measured <0.65 790 ± 23 2.80 ± 0.13 77.9 ± 2.3 manage the data generated is of critical importance in order to
Indicated 0.65 - 1.1 1734 ± 87 2.84 ± 0.17 174.0 ± 14.0 implement the suggested approaches.
Inferred >1.1 103 ± 9.4 2.92 ± 0.16 10.5 ± 1.9
A LINK TO PRODUCTION SCHEDULING AND
SOME RELATED ASPECTS
important difference from traditional classification methods in
The approaches and examples discussed in the preceding sections
this example is that the classification is independent of the block
do not consider the potential influence of production scheduling to
average grade.
ore reserve classification. It can be argued that sequencing and the
Please note that a key expectation of resource classification is time of mining adds a new dimension to decisions about reserve
to support mine planning. Parker’s approach (Harry Parker, classification. This is because the criteria for deciding what is an
personal communication) commonly used, (shown also by Lopes acceptable risk in one year may not necessarily be the same as that
et al, 2010, in this volume) suggests that Measured should in for decisions ten years later and the risk associated with mining
support three months production increments and Indicated part of an orebody this year should not necessarily be assessed in
annual production increments. Although at the time of this study
the same way as what is to be mined some years later when
this was not implemented, the approach presented can easily and
information and conditions will be different.
should be implemented, so as to state confidence levels in terms
of production increments used for planning. An extension of the example from the alluvial diamond deposit
presented earlier is used here to illustrate the quantification of
uncertainty in reserves (and thus meeting production targets) –
Selecting thresholds at a copper deposit here the total volume of ore material mined. First, a production
schedule is developed through conventional means for twelve 500
The example of a porphyry copper deposit in Canada further × 500 m2 blocks and considers mining four such blocks per period
illustrates the use of a CCV based classification for resource/ over three production periods. Then, uncertainty assessment is
reserve classification. The example considers the part of the performed from a set of simulated orebody models at the same
deposit within the optimal pit limits (see Leite and block support-scale to model the uncertainty in the volume for
Dimitrakopoulos, 2010, in this volume). For the present study, each production period. Figure 10 shows the ccdfs for volumes of
25 realisations of the copper grades of 20 × 20 × 10 m3 blocks are material mined in each of the three production periods, thus
used. Figure 8 shows tonnage of ore above the 0.3 per cent Cu describing their uncertainty. In addition, Figure 10 shows the risk
cut-off used here plotted against the CCV for the 25 simulated profile of the expected volume for three production periods. If risk
realisations of the copper orebody generated directly at the block is a main criterion in reserve classification, then the reserve in the
size required (eg Benndorf and Dimitrakopoulos, 2007). Similarly first mining period should perhaps be classified in a different
to the previous case study, the figure shows a distinctly increasing category from that in the third period.
120
100
Tonnage above cut-off (tonnes x 106)
80
60
40
20
0
0 0.2 0.4 0.6 0.8 1 1.2 1.4
CCV
FIG 8 - Tonnage of ore versus the CCV for 25 realisations of the copper deposit.
desired blending of materials within periods that provide overall

prespecified risk profile in a scheduling period, thus optimising the
classification of reserves. It is possible to substantially improve
and provide more realistic assessments and classification of
reserves when schedules and resource uncertainty are jointly
managed.
CONCLUSIONS
Mineral resources and ore reserves are classified as Measured,
Indicated, Inferred and Proven or Probable respectively both for
public reporting and internal company assessments. Resource/
reserve classification is traditionally based on geometric criteria,
such as the number of samples or drill holes within a volume, or
some form of estimation variances. Although transparent and
relatively easy to understand, these criteria ignore local variability
of grades within a deposit and can lead to oversimplification or
misrepresentation of the confidence of estimated grades.
In this paper, an approach integrating both geometric criteria
FIG 9 - Blocks in a section of a gold load classified using the CCV and in situ grade variability is described and its application
classification (left) and using available drilling (right). explored. The approach is based on conditional simulation and
risk quantification in five steps: generation of suitable
The integration and understanding of the interaction between realisations of the deposit at the appropriate block-support scale,
reserve classification with the mining sequence and production derivation of conditional cumulative distribution functions of
scheduling are traditionally not considered, although in the past, block grades or attribute selected, utilisation of the Conditional
conditional simulation has been used to test uncertainty of Coefficient of Variation (CCV) of the grade ccdf within a block
production schedules (Ravenscroft 1992; and others). However, it as the uncertainty measure for resource classification of each
is logical, as the simple example above suggests, to consider block in the orebody model, classification of resources/reserves,
enhancing the classification approaches. For example, the and evaluation and production of risk profiles for estimates.
uncertainty measures previously described may be used to classify The practical aspects and intricacies of the proposed approach
resources/reserves given a production schedule and consider are explored in three applications at an alluvial diamond deposit,
different classification criteria for the mining periods so as to an epithermal gold lode and a porphyry copper deposit. An
reflect timing. interesting observation, common to these applications, is that there
It is also and perhaps more important to stress that scheduling are natural groupings of the CCV of mining blocks that allow for
can be performed under geological uncertainty and geological risk rational categorisation of resources or reserves in a transparent and
discounting can be applied within the scheduling formulations. technically defensible way. With this tool, along with existing
This discounting concept is introduced in Dimitrakopoulos and practices, such as examining cross-sections of the deposit, to
Ramazan (2004) and used in new stochastic long-term mine ensure the definitions in the National Reporting Codes are met. As
production schedulers (Ramazan and Dimitrakopoulos, 2009; also, a result, reporting of resource/reserve classification and may
Benndorf and Dimitrakopoulos, 2010, in this volume) leads to a support better investment decision making.
FIG 10 - Conditional cumulative distribution functions describing the uncertainty on the volume of mineralised materials for the production
periods (left) and risk profile on the expected volume for three production periods (right). Variances plotted are not cumulative.
A key improvement to this approach could be linking it to Froidevaux, R, Roscoe, W E and Valiant, R I, 1986. Estimating and
mine production scheduling, first, so that classifications account classifying gold reserves at Page-Williams C-zone: A case study in
for production units over time – what is to be mined in ten years non parametric geostatistics, in Ore Reserve Estimation: Methods,
should not perhaps be classified as what is to be mined next year Models and Reality (eds: M David, R Froidevaux, A J Sinclair and
M Vallée), pp 280-300 (Canadian Institute of Mining: Montreal).
and secondly, to consider control of the classification of
resources/reserves through risk blending and deferment, a Giroux, G H, Sinclair, A J and Miller, J H L, 1986. Production quality
capability that stochastic schedulers can offer. control experiments, in Ore Reserve Estimation: Methods, Models
and Reality (eds: M David, R Froidevaux, A M Sinclair and
M Vallée), pp 238-260 (Canadian Institute of Mining: Montreal).
ACKNOWLEDGEMENTS Joint Ore Reserves Committee (JORC), 2004. Australasian Code for
Thanks are in order to the industry members of the COSMO Reporting of Exploration Results, Mineral Resources and Ore
reserves (JORC Code).
Laboratory: AngloGold Ashanti, Barrick, BHP Billiton, De
Beers, Newmont, Vale and Vale Inco, as well as NSERC Journel, A G and Kyriakidis, P C, 2004. Evaluation of Mineral Reserves
– A Simulation Approach, 232 p (Oxford University Press: New York).
(Discovery Grant 239019), the Canada Research Chairs Program
and CFI are gratefully acknowledged. Leite, A and Dimitrakopoulos, R, 2010. Production scheduling under
programming at an open pit copper mine and comparison to
REFERENCES conventional scheduling, in Advances in Orebody Modelling and
Arik, A, 2002. Comparison of resource classification methodologies with Strategic Mine Planning I (ed: R Dimitrakopoulos), pp 35-40 (The
a new approach, in Proceedings Applications of Computers and Australasian Institute of Mining and Metallurgy: Melbourne).
Operations research in Mineral Industry (APCOM) 2002. Lopes, J A, Rosas, C F, Fernandes, J B and Vanzela, G A, 2010.
Audet, M and Ross, A F, 2007. Koniambo lateritic Ni-Co deposits, New A geology-based joint conditional simulation of a lateritic nickel
Caledonia – A case study from geological modelling to mineral deposit – Quantification of risk in grade tonnage curves and resource
resource classification, in Orebody Modelling and Strategic Mine categorisation, in Advances in Orebody Modelling and Strategic
Planning, second edition (ed: R Dimitrakopoulos), pp 235-244 (The Mine Planning I (ed: R Dimitrakopoulos), pp 205-214 (The
Australasian Institute of Mining and Metallurgy: Melbourne). Australasian Institute of Mining and Metallurgy: Melbourne).
Benndorf, J and Dimitrakopoulos, R, 2007. New efficient methods for Murphy, M, Parker, H, Ross, A and Audet, M, 2004. Ore-thickness and
conditional simulation of large orebodies, in Orebody Modelling and nickel grade resource confidence at the Koniambo nickel laterite
Strategic Mine Planning, second edition (ed: R Dimitrakopoulos), deposit in New Caledonia – A conditional simulation voyage of
pp 61-68 (The Australasian Institute of Mining and Metallurgy: discovery, in Geostatistics Banff 2004 (eds: O Leuangthong and
Melbourne). C Deutsch) (Springer: Dordrecht).
Benndorf, J and Dimitrakopoulos, R, 2010. Stochastic long-term Parker, H. Personal communication.
production scheduling of iron ore deposits: Integrating joint Ramazan, S and Dimitrakopoulos, R, 2009. Production scheduling with
multi-element geological uncertainty, in Advances in Orebody uncertain supply: A new solution to the open pit mining problem,
Modelling and Strategic Mine Planning I (ed: R Dimitrakopoulos), European Journal of Operations Research (submitted) and COSMO
pp 151-158 (The Australasian Institute of Mining and Metallurgy: Report 2007, pp 257-294.
Melbourne). Ravenscroft, P J, 1992. Risk analysis for mine scheduling by conditional
Canadian Securities Commissions (CSA), 2005. National instrument simulation, Transactions of the Institutions of Mining and
43-101: Standards for disclosure for mineral projects. Metallurgy, Mining Technology, 101:A104-A108.
David, M, 1977. Geostatistical Ore Reserve Estimation, p 364 (Elsevier: Sabourin, R L, 1983. Application of Geostatistical method to
The Netherlands). quantitatively define various categories of resources, in Geostatistics
David, M, 1988. Handbook of Applied Advanced Geostatistical Ore for Natural Resources Characterisation Part 1, series C:
Reserve Estimation, p 217 (Elsevier: The Netherlands). Mathematical and Physical Sciences (ed: G Verly) 122:201-216.
David, M and Dagbert, M, 1977. Predicting vanishing tons before Sinclair, A J and Blackwell, G H, 2002. Applied Mineral Inventory
production starts, or small blocks are not good for planning in Estimation, 381 p (Cambridge University Press: New York).
porphyry deposits, paper presented to annual meeting of the American South African Mineral Resource Committee (SAMREC), 2000. South
Institute of Mining, Metalurgy and Petroleum Engineering, Atlanta. African reporting of mineral resources and mineral reserve, Southern
Dimitrakopoulos, R and Ramazan, S, 2004. Uncertainty based production African Institute of Mining and Metallurgy.
scheduling in open pit mining, SME Transactions, 316:106-112. United States Securities and Exchange Commission (SEC), 2006.
Dohm, C, 2005. Quantifiable mineral resource classification – A logical Industry Guide 7.
approach, in Geostatistics Banff 2004 (eds: O Leuangthong and Wawruch, T, Jorge, F and Betzhold, J F, 2005. Mineral Resource
C Deutsch), 1:333-342 (Springer: Dordrecht). classification through conditional simulation, in Geostatistics Banff
Duggan, S and Dimitrakopoulos, R, 2005. Application of conditional 2004 (eds: O Leuangthong and C Deutsch), 1:479-489 (Springer:
simulation to quantify uncertainty and to classify a diamond Dordrecht).
deflation deposit, in Geostatistics Banff 2004 (eds: O Leuangthong
and C Deutsch), 2:419-428 (Springer: Dordrecht).
HOME
Reducing the Geological Risk in Mining an Orebody by Using

Borehole Radar for Strategic Mine Planning
C Kemp1, P Du Pisani2, A Bray3 and G Chitiyo4
ABSTRACT detailed information is needed to negotiate geological

discontinuities on the panel scale. The work described in this
Borehole radar is a proven geophysical technology that can be used to
reduce the geological risk in extracting an orebody and therefore improve
paper shows how mining operations can be supported by tactical
mine planning. This paper will present a case study, where borehole radar underground Borehole Radar (BHR) surveys, and puts forward
was used within the mining cycle to map out orebody blocks to assist with the proposition that BHR can be used to better manage many of
the mine design months ahead of mining. Refined equipment and the risks that are inherent in concentrating production on a small
procedures now enable a slim-line borehole radar tool to be deployed on number of high output mechanised panels.
the end of drill rods, making it easier and faster to deploy in an
underground mine with minimal impact on the mining process. The case BOREHOLE RADAR FOR STRATEGIC MINE
study is from a South African platinum mine, where borehole radar was
used successfully in identifying small-scale disruptions in the platinum reef PLANNING OF PLANAR OREBODIES
that may influence the way in which it will be mined. Slim-line borehole BHR is an electromagnetic subsurface imaging technique
radar tools are deployed in applications where boreholes are drilled parallel
designed for imaging or detecting discontinuities in restive hard
or subparallel to the orebodies in order to delineate them and any disruptive
structures on them, which change the normal orebody location, thickness
rock formations. Radar works best when the tool is deployed in a
and sometimes composition. Advance knowledge of how the orebodies are resistive host rock and used to image a conductive orebody.
displaced impacts on how the orebody is mined. Mine planning can then be Reflections are normally caused by sharp changes in the
adapted to mine the orebody more economically and potentially hazardous electrical properties of the rock, and can yield information on
situations can be negated. geological change, faults/fractures, voids, dykes, orebody
geometry, etc. On the other hand, conductive or highly fractured
rock formations increase the attenuation (scattering) of
INTRODUCTION electromagnetic energy, making BHR less effective.
Globally, the mining industry and particularly Anglo Platinum are
adapting to rapidly rising labour costs and improvements in Instrumentation
technology. It is moving from a tradition in which small, flexible
airleg mining crews advance a large number of ~30 m wide A BHR tool consists of a transmitter and receiver, which can be
metre-high panels about 17 m per month, through concentrated combined in one antenna, resulting in a ‘single-stick’ probe, or as
capital intensive trackless mining operations advancing at ~50 m two separate antennas in two separate probes, usually separated by
per month on a front five times wider, to experimental continuous a fibre optic spacer. Both antenna configurations have their
mining operations that are projected to achieve advance rates of advantages and disadvantages, however due to logistical and space
over 100 m per month. Negotiating geological discontinuities constraints in underground hard rock mining, usually the
interferes with production to a degree that generally rises with the ‘single-stick’ probes are used. The current BHR design has come a
level of mechanisation and the rate of face advance. For example, long way in the last couple of years. Improving from probes that
in platinum reef mining, potholes cause the most frequent were connected by fibre optics to a digital acquisition system at
disruptions. A faster advance rate increases the frequency of the borehole collar (Vogt, 2006), to the single-stick radars with
exposed potholes – the greater the capital concentration, the more onboard memory (Bray et al, 2007) enabling surveying to be
serious the impact upon ore production. The more unexpected the conducted on the drill rods directly after the boreholes are drilled,
pothole, the lower the stockpiles and thus the ripple of feedstock in the same manner that borehole deviation surveys are conducted.
that enters the mill is greater. The latter technique now enables BHR to be surveyed in all
underground boreholes, where previously broken ground and
Observations such as these have stimulated recent geophysical
difficult end of hole pulley deployment methods hindered its
reconnaissance efforts across a number of different commodities.
success. The slim-line BHR tool has been designed specifically for
Three-dimensional surface seismic surveys costing around
AU$250 000 per square kilometre can map faults of several underground mining. Its diameter enables it to fit into exploratory
metres of throw, and most shallow potholes greater than 50 m geological or cover drill holes with diameters smaller than 48 mm.
wide in plan on those parts of platinum reef that are neither The BHR tool is broadband, with a centre frequency of 50 MHz.
overshadowed by old workings on an overlying platinum reef, This equates to a range in most hard-rock environments of at least
nor vulnerable to screening by seismic reverberation between 30 m from the borehole, but up to 80 m in some rock types. The
horizons that lie either above or near the platinum reef. This dielectric properties of the host-rock determine the range that
technique is useful on a broader exploration scale, but more BHR can achieve. Slim-line BHR tools are omni-directional.
Reflection data is recorded from signals received from 360°
surrounding the boreholes. While a lot of research is being
1. Business Development Manager and Senior Geophysicist, Geomole conducted in directional radar technology, it is difficult to fit an
Pty Ltd, Australian Technology Park, International Business Centre, efficient directional radar antenna in a probe that will fit in
Suite 108, 2 Cornwallis Street, Eveleigh NSW 2015. underground mining drill holes, eg less than 48 mm diameter.
2. Manager, In-Mine Geophysics, Anglo Technical Division, Borehole radar survey design
Geosciences Research Group, Office 553, 45 Main Street,
Marshalltown Johannesburg 2107, South Africa.
The borehole design and layout is absolutely critical in successful
3. SAusIMM, PhD Student, The School of Geosciences, Madsen
radar implementation. For thin reef or vein style deposits, BHR is
Building (F09), The University of Sydney, Sydney NSW 2006.
generally conducted in a series of regularly spaced single-borehole
4. Chief Geologist, Anglo Platinum, Rustenburg Section, South Africa. reflection surveys, however, the borehole design depends greatly
C KEMP et al
on the type of orebody and the geological discontinuities one of a parallel set of nadir lines is (Davis, 1986) directly
associated with it. Here we will look at a BHR survey design for proportional to the ratio of the pothole’s periphery and the nadir
thin reef or vein style deposits. line spacing D:
Survey design is conditioned by limitations of site access and
2( a 2 + b 2 )
possible parallel layering due to bedding or joint planes within P=
the host rock. Consider, in Figure 1, the descent of a BHR D2
through a succession of bedding planes, the UG2 and UG1.
The probability of detecting a pothole during a BHR profile of
Passage through each plane in turn generates an inverted ‘V’
a borehole is of course higher, for on the one hand many potholes
pattern in the time-section of a trace gather. Echoes from off-axis
have deep roots and generate sideswipe, and on the other hand,
diffractors appear as inverted hyperbolas. Similar hyperbolas
even if the pothole does not lie on a nadir line, it can and will
appear on surface seismic profiles and on raw Synthetic Aperture
influence the first arrival if it lies within the first Fresnel zone,
Radar (SAR) trace gathers. The BHR-SAR analogy is highly
shown in Figure 2a. Nadir lines form a framework that may be
useful, but there are very significant differences between the two.
used to reconstruct a 3D model of a buried surface, Figure 2b.
One lies in the fact that targets of interest in BHR may lie on
Therefore the borehole design for BHR should consider the size
many planes. In SAR, most targets lie on the ground. Another is
of the object that is required to be detected and the range to the
that directional antennas allow SAR systems to eliminate ground
target plane.
echoes or nadir reflections and to concentrate on imaging the
backscatter or sideswipe from off-axis targets. In hard-rock BHR,
it is nearly impossible to suppress ground echoes or nadir CASE STUDY – PLATINUM REEF MINING
reflections, because efficient directional VHF-UHF antennas
cannot fit into narrow (47 mm) boreholes. BHR time sections
tend to be dominated by specular reflections from areas that Overview
straddle nadir lines. Nadir lines are the loci of reflection points The 2060 Ma old Bushveld Complex is a massive layered igneous
on a reflecting plane. The normals to the surface at these points intrusion, with outcrop extremities of ~450 km east-west and
collectively define sagittal planes. ~300 km north-south. It has variously been attributed to the
If transmitter and receiver are at identical heights above a plane, subduction of a mid-oceanic spreading centre, a multiple meteorite
the sagittal plane will pass through the bisector. If there are many impact, and a mantle plume. It can be compared to an enormous,
parallel boreholes, BHR surveys may paint near-parallel nadir irregularly-shaped saucer with its centre deeply buried, but its rim
lines on the bedding planes. The probability P of intersecting an exposed. The saucer is made up of distinct thin layers. Two of
elliptical pothole or similar disruption, semi-axes a and b, with these layers, the Merensky and the underlying UG2 chromitite
(A) (B)
FIG 1 - (A) Example borehole radar layout and (B) resulting borehole radar time section.
(A) (B)
FIG 2 - (A) A pothole within the first Fresnel zone surrounding the nadir line will influence the first arrive, and (B) a small elliptical
pothole lies on a surface generated from a framework of nadir lines.
REDUCING THE GEOLOGICAL RISK IN MINING AN OREBODY BY USING BOREHOLE RADAR FOR STRATEGIC MINE PLANNING
reef, are members of a well layered ultramafic to mafic succession

known as the Rustenburg Layered Suite. These two reefs contain
well over half of the world’s platinum reserves (Viljoen and
Schurmann, 1998; Cawthorn, 1999). The reefs are generally
shallow dipping tabular orebodies. They have been tracked using
exploratory drilling and seismic methods over kilometres within
the Bushveld complex (Eales et al, 1993). On a mining scale, the
reefs can be disrupted by various smaller-scale features such as
dykes, faults, slumps (called potholes), rolls and Iron-Rich
Ultramafic Pegmatites (IRUPs). Predicting the location of these
disruptions prior to mining affects resource calculation, mine
planning, the extraction of the orebody, as well as mine safety (Du
Pisani et al, 2008).
BHR has been used successfully in many of South Africa’s
underground platinum mines to delineate both of the the Merensky
and UG2 platinum reefs (Van Schoor, Du Pisani and Vogt, 2006;
De Vries and Du Pisani, 2007). Here we will consider potholes
and rolls that disrupt the reef. According to Lomberg et al (1999),
potholes are more disruptive to layering and mining than rolls. FIG 3 - Schematic of borehole radar survey design as seen in
Mining operations near potholes are disrupted by the sudden dip in profile (after Du Pisani et al, 2008).
the strata, increasing the number of joints and talc-filled fractures,
cross cutting pegmatite veins and a weakening of the roof. The and then shot using BHRs that moved up each hole in turn at
reef thins gradually on approach to a shallow conformable 10 m/min.
pothole. More rapid reef thinning indicates an approach to a deep
unconformable pothole.
Borehole radar results and interpretation
At the edge the upper leader chromitite suddenly
breaks through the lower leader chromitite The five panels in Figure 5 show the resulting trace gathers. It is
layers and then, over two to 5 m, cuts sharply difficult to identify the different reflections in the data from the
into the underlying feldspathic pyroxenite … Borehole Radar data alone. This is difficult because slimline
sometimes … so sharply that it cuts back BHRs have toroidal radiation patterns and therefore collect data
underneath the pothole reef (after Lomberg et al, from 360° around the drillhole. Distinguishing between arrivals
1999). from planes above the borehole and from those below is
Potholes range from one to 500 m in diameter, drop from one to accomplished by modelling. The modelling procedure is
100 m into the footwall and can have sloping, steep, or even illustrated in Figure 6, where four model boreholes track east-west
overhanging margins. They can appear as simple washouts, as and a model tie hole tracks north-south. The inclination of the
slumps, or as deeply sunken cones, the sides of which may be model mirror plane above the boreholes can be adjusted, as can
the plane’s elevation. On the upper-right of Figure 6, the mirror
decorated by shards of UG2 chromitite and into which are sucked
appears as a horizontal line, while the borehole appears as a curve.
substantial amounts of the hanging wall. The layers above a deep
If the mirror plane’s inclination matches that of the bedding
pothole are frequently deformed, sometimes plastically. Brittle
planes, then as its elevation rises, the mirror move-out line will
failures above some potholes fault the hanging wall and hamper successively match the dashed arrivals first from a model footwall
efforts to mine into or through them. UG1 and then from a model hanging wall UG2.
Figure 7 illustrates the practical application of mirror plane
Borehole radar survey design synthesis. There it is shown in close-up time sections from the
BHR test surveys were conducted in 2003 at Anglo Platinum’s three nearly parallel boreholes Number 06, Number 07 and
Bleskop shaft that is close to Rustenburg (South Africa) in order Number 08. These were drilled westwards from stations in a
to detect disruptions in the UG2 reef, demonstrate that BHR can horizontal north-south cross-cut. Their collars were spaced at 35 m
be used to optimise mine layouts and to test the viability of and 25 m. Smooth red curves, superimposed on the time sections,
applying BHR in underground platinum mines. The locations of show the model mirror-plane moveouts at elevations of 778.5,
boreholes drilled for the application of BHR are limited by the 805.5, 808.8, 812.8 and 818.3 (measured normal to bedding). The
access available from mining excavations. Platinum reefs in lowest synthetic, at 27.5 m beneath the next – the Bleskop Marker
South Africa are traditionally mined using a conventional breast (BkM) – tracks echoes from the UG1. Echoes from the UG2 (at
stoping mining method. Radar boreholes are usually drilled 808.8 m nominal) indicate a separation of 3.3 m from the BkM.
below the reef in the footwall, either from haulages or crosscuts The echo 4 m higher (at 812.4 m nominal) is linked to a group of
and are oriented parallel or subparallel to the reef plane – see beds known as the triplets. And finally, the top echoes, tracked at
Figure 3. The area covered by radar boreholes is usually the size 818.3 m, are associated with the Pyroxenite-Anorthosite (PxA)
of a planned stope, around 200 × 200 m. This enables mapping boundary in the UG2’s hanging wall. Each of the sections can be
the reef topography prior to mining. examined in detail and converted into a section interpretation. The
Access to the UG2 at Bleskop was established in the section interpretation for borehole Number 8 is shown in Figure 8.
mid-nineties, through a network of strike haulages that were It shows the UG2 breaking away from the normal stratigraphy at
driven ten to 20 m below the reef. Four 180 m long AXT the beginning and end of the hole. Shading to the right indicates
boreholes, numbered 6, 7, 8 and 9 in the perspective view in that certainty drops when the borehole, shown in red, ascends into
Figure 4, were drilled west from stations spaced at ~35 m in a a perturbed and thinned section of the UG2 reef. The pothole on
north-south stub-end that ran out ~10 m beneath a moderately the left is approximately 20 m deep and the ‘mine-able corridor’ is
large pothole in the UG2 from an east-west strike haulage. A fifth 50 m wide. Slumping, reef thinning and shearing mark the western
~200 m long tie borehole, Number 10, was drilled south from the boundary of this corridor. From this section it is easy to identify
haulage. The holes cover a mining block spanning approximately that the centre of the section would be the only suitable place for
20 000 m2. They were rigged with pulleys, surveyed accurately undisrupted mining of the UG2 platinum reef.
C KEMP et al
FIG 4 - Perspective view of the radar boreholes drilled at Bleskop.
FIG 5 - Borehole radar gathers, band-pass filtered and gain-swept. Time runs vertically downwards. Distance from hole collars runs
horizontally across the sections. Curved lines track mirror reflections from planes 37.8 m apart.
FIG 6 - Identification of arrivals from specific bedding planes using modelling. Though individual layers may differ locally in thickness, the
stratigraphy, summarised by colours marked both on the bedding plane normal and on the sagittal plane upper right, is known from
geological logs.
FIG 7 - A practical application of mirror plane synthesis: matching time sections from boreholes #06, #07 and #08 to a 3D plane-parallel
layered model of the UG2 reef and its surrounding host rocks. Thin red lines indicate layered stratigraphy. Dotted lines indicate deviations
from this layered stratigraphy.
C KEMP et al
FIG 8 - Section interpretation of the sagittal plane of Borehole Number 08, based on analyses of the sections presented in Figure 5
and Figure 7.
Strata that are healthy from a mechanised mining point of view isometric view of the potholed reef, Figure 10 shows the
usually match layer cake time-of-flight models very well. geological interpretation in plan and Figure 11 shows BHR
Deviations from this layer cake stratigraphy indicate areas of derived contours both of the reef and the hanging wall PxA.
concern for mining, as disruptions such as potholes have disturbed
the reef in these locations. The resulting interpretation of the Discussion and impact on mining
stratigraphy for this mining area is illustrated in Figure 8.
Evidence of disturbance comes from the 2D radar time sections. The results of the BHR indicated various areas of concern for
The identification of 3D shape comes from interpolating between mining. However, these BHR trials at Bleskop were conducted
arrivals. Recovery of sagittal planes which are consistent with the when the technique was still considered to be experimental.
interpolated 3D surface can be achieved by iteration. Figure 9 is an Therefore the BHR results were not received in time or with
FIG 9 - Three-dimensional BHR image of a large spoon-shaped UG2 pothole.
FIG 10 - Working interpretation of the surfaces contoured in Figure 9.
FIG 11 - Contours at 2 m intervals link the hanging wall pyroxenite – anorthosite (upper) to the underlying UG2 reef in a potholed panel at
Bleskop Platinum Mine. A large volume of hanging wall material appears to have been drawn laterally into the pothole upper right.
enough weight to influence the mine design and avoidable mining TABLE 1
and development took place. This trial provided a good test case to Financial impact summarised (after Du Pisani et al, 2008).
show how much could have been saved on unnecessary mining
and development when using BHR. Subsequent mining of this Item Extent Tonnes Cost
mining block (green panels in Figure 12) showed that the Avoidable mining 4300 m2 14 400 US$1.03 M
conclusions drawn from the BHR results were true and that the
Avoidable development 338 m US$230 k
large anomaly identified on the western side of the block was a
significant reef roll. Figure 13 illustrates the unnecessary mining Total US$1.26 M
and development that took place in this mining area at Bleskop.
The costs that could have been saved within the Bleskop African platinum mines. It should also be noted that not all
pothole, if the BHR results had been used, are summarised in panels lend themselves readily to high-levels of mechanisation.
Table 1 (after Du Pisani, 2008). All costs used are based on Many contain unsuspected faults that may delay or halt
November 2007 Platinum-Group Metals (PGM) prices as well as production. BHR surveys can add significantly to the value of an
generalised costs for mine development and drilling at South orebody not only by enabling optimum equipment to be selected
C KEMP et al
FIG 12 - Geological features encountered during subsequent mining of Bleskop mining block (after du Pisani et al, 2008).
FIG 13 - Unnecessary mining and development that took place within the Bleskop pothole (after Du Pisani et al, 2008).
for a given task, but also by allowing the effort put into the footwall that involves one kilometre of boreholes would cost less
development of different faces to be synchronised so that they do than ten to 20 m of unnecessary development.
not all go down together. BHR surveying adds to predevelopment
costs. A kilometre of AXT drilling in the footwall takes about SUMMARY AND CONCLUSIONS
one hundred pneumatic crew shifts and costs about AU$50 000. The case study presented in this paper shows that BHR, if applied
Boreholes are 50 to 100 times as expensive as BHR. This means correctly, can significantly contribute to lowering the costs of
that the BHR surveying of a 200 m × 200 m stope from the operating a platinum mine. Strategically conducting BHR surveys
both prior to and within the mining cycle improves the geological Cawthorn, R G, 1999. The platinum and palladium resources of the
knowledge about the orebody. BHR surveys enable strategic mine Bushveld Complex, South African Journal of Science, 95:481-489.
planning for the appropriate mine design and the efficient Davis, J C, 1986. Statistics and Data Analysis in Geology, second edition,
deployment of mining teams. But most of all, BHR can identify p 290 (John Wiley: New York).
additional resources in structurally complex areas of the mine De Vries, P and Du Pisani, P, 2005. Borehole radar delineation of the
and sterilise areas that are too disrupted to mine. At Anglo UG2 reef at Modikwa Mine, in Proceedings Ninth SAGA Biennial
Platinum, BHR has since moved beyond the realm of trial Technical Meeting and Exhibition (South African Geophysical
surveys and is being routinely applied as a geological mapping Association).
tool at various mines. Du Pisani, P, Coomber, S, Chitiyo, G, Daniso, V, Mampa, S and Mason, I,
2008. Using borehole radar to detect disruptions in platinum reefs in
South Africa – The financial implication, in Proceedings 12th
ACKNOWLEDGEMENTS International Conference on Ground Penetrating Radar, 16-19 June,
Birmingham, UK.
The authors thank Anglo Platinum for allowing these results to be
published. In particular, Theo Pegram, Gordon Chunnett and Du Pisani, P, Mason, I M and Daniso, L N, 2007. Borehole radar
delineation of the Brakspruit regional pothole at Anglo Platinum’s
Marshall Patterson are thanked for their continued support of
Rustenburg Section, in Proceedings Tenth SAGA Biennial Technical
borehole radar as a delineation tool for platinum reefs. Tony Meeting and Exhibition, (South African Geophysical Association).
Redman (Anglo American), Charles Pretorius (ATD) and Alan
Eales, H V, Botha, W J, Hattingh, P J, De Klerk, W J, Maier, W D and
King (ATD) are acknowledged for their support, as are all the Odgers, A T R, 1993. The mafic rocks of the Bushveld Complex: A
Rustenburg geologists, borehole radar surveyors and drilling review of emplacement and crystallisation history and
contractors who have made these borehole radar surveys possible. mineralisation, in light of recent data, Journal of African Earth
The authors would like to acknowledge staff members and miners Science, 16:121-142.
of RPM Rustenburg Section for their vision and support of the Lomberg, K G, Martin, E S, Patterson, M A and Venter, J E, 1999. The
implementation of geophysical programmes. The professional morphology of potholes in the UG2 Chromitite Layer and Merensky
enthusiasm of Francois Vos, Vasek Novi, Kabelo Thlapi and Luke Reef (pothole reef facies) at Union Section, Rustenburg Platinum
Zindi made this study both possible and enjoyable. We thank the Mines, South African Journal of Geology, 102(3):209-220.
drillers of Rosond and all of the miners at Brakspruit, Bleskop, Van Schoor, M, Du Pisani, P and Vogt, D, 2006. High-resolution,
Waterval and Boschfontein for their competence and invariably short-range, in-mine geophysical techniques for the delineation of
kind assistance. The authors gratefully acknowledge the scientific South African orebodies, South African Journal of Science,
and technical contributions of Iain Mason, Jonathan Hargreaves, 102:355-360.
Binzhong Zhou, Lang Hames, Brian Woods, Wessel van Brakel Viljoen, M J and Schürmann, L W, 1998. The Mineral Resources of South
and Paul van der Merwe. Africa (Council for Geoscience: Pretoria).
Vogt, D, 2006. A borehole radar system for South African gold and
REFERENCES platinum mines, South African Journal of Geology, 109:521-528.
Bray, A, Sindle, T, Mason, I M, Palmer, K, Cloete, J, Steenkamp, J, Du

Pisani, P and Trofimczyk, K, 2007. Using slimline borehole radar
from cover holes, in Proceedings Tenth SAGA Biennial Technical
Meeting and Exhibition (South African Geophysical Association).
HOME
Strategic Optimisation of a Vertical Hoisting Shaft in the Callie

Underground Mine
M G Volz1, M Brazil2 and D A Thomas3
ABSTRACT
The Callie underground mine, located in the Tanami Desert in the
Northern Territory, includes two parallel declines accessing a large
orebody extending some two kilometres below the surface. One of several
ideas considered in strategic mine planning is to incorporate a vertical
hoisting shaft and an orepass as an alternative to trucking material to the
surface along the declines. In this work, we use network optimisation
techniques to investigate the feasibility of the proposed system, and to
mathematically determine the optimum positions and geometry of the
shaft, orepass and surrounding infrastructure. We propose a modelling
procedure taking aspects from a mathematical problem, called the Fermat-
Weber problem, which asks for a point minimising the sum of weighted
distances to a given set of points. We describe the implementation of the
procedure into a computer program for solving the problem iteratively, and
present results over a range of infrastructure and haulage costs, decline
gradients and life-of-mine (LOM) schedules.
INTRODUCTION
Located in the Tanami Desert in the Northern Territory,
approximately 550 km north-west of Alice Springs, the Tanami
operations comprise a processing facility, several open pit mines, FIG 1 - Typical cross-section through the orebody (looking west).
the Tanami Mill and the Callie Underground Mine. Gold was
discovered at Tanami in 1900 and modern mining began in 1983
following an agreement with traditional landowners. Initial
production came from the open pit mines at the Granites, while
current production comes from the high-grade Callie underground
mine at Dead Bullock Soak. The Callie underground mine
services a large orebody running approximately in an east-west
direction, plunging into the ground at an angle of about 45°
towards the east (Figure 1). The orebody is divided into two major
veins. At the time of the study, the orebody had been accessed by
a single decline, called the Callie decline, with material having
been mined to a depth of about 1000 m below the surface
(Figure 2). Rock is extracted from the orebody in primary stopes,
which are replaced with fill material to allow adjacent secondary
stopes to be subsequently extracted. Once loaded onto trucks, ore
is hauled along cross-cuts (horizontal tunnels) to the decline.
Levels are at 40 m vertical intervals. In addition to the planned
extension of the primary Callie decline, a secondary decline,
called the Wilson Drill Decline (WDD), is to branch out from the
FIG 2 - The Callie underground mine (looking north), including the
Callie decline about 950 m below the surface (Figure 2). Both
existing Callie decline (white), its planned extension (red/blue) and
declines are to have fixed gradients. Levels servicing the WDD are
the proposed Wilson drill decline (pink).
also at 40 m vertical intervals, however they are offset from the
Callie levels. The Callie decline services the Wilson shoot, while
the WDD is to service a second shoot called the Federation shoot. hoisted to the surface via the shaft. This method can provide
One of several ideas considered in strategic mine planning is significant reductions in operating costs, although it requires a
to incorporate a vertical hoisting shaft and an orepass as an large capital cost associated with a hoisting shaft. In addition to
alternative to trucking material to the surface along the declines. the shaft, it was proposed to include an orepass into the mine. An
Using this system, ore is hauled to a common tipping level, orepass is a near-vertical chute down which ore from upper
called the haul level, where it is crushed, loaded into a skip and levels is dropped to the haul level (which is three levels above the
shaft base), and transported to the shaft in one of several ways,
for example:
1. Analyst, TSG Consulting, Level 11, 350 Collins Street, Melbourne
Vic 3000. Email: [email protected]
• Ore is loaded into a truck by a load-haul-dump vehicle at the
bottom of the orepass. It is then trucked from the bottom of
2. Senior Lecturer, Department of Electrical and Electronic the orepass to the shaft.
Engineering, The University of Melbourne, Parkville Vic 3010.
Email: [email protected] • A load-haul-dump vehicle trams ore directly from the base of
the orepass to the shaft.
3. Professor and Head, Department of Mechanical and Manufacturing
Engineering, The University of Melbourne, Parkville Vic 3010. • If the horizontal distance between the orepass and the shaft is
Email: [email protected] greater than say 300 m, a loading chute may be installed at the
M G VOLZ, M BRAZIL and D A THOMAS
base of the orepass, allowing trucks to be loaded automatically

before transporting ore to the shaft. For operational reasons the TABLE 2
orepass is assumed to be constrained to the northing 9250N, Base, probable and best life-of-mine schedules.
which is about halfway between the two declines in plan. Callie decline Wilson drill decline
The primary goal of this investigation is to mathematically Rate Quantity Final Rate Quantity Final
determine: Name
(Mt/a) (Mt) year (Mt/a) (Mt) year
• the optimum position (depth and plan coordinates) of the Base 2 10 2015 0.35 1.4 2014
hoisting shaft; Probable 2.5 17.5 2018 0.5 2.5 2015
• the optimum position (top, bottom and plan coordinates) of Best 3.5 30 2019 0.5 3.5 2017
the orepass and identification of the levels which access the
orepass; and
to avoid faults by at least 50 m. The main haulage drive is
• the optimum geometry of the main haulage drive network at allowed to pass through a fault so long as it does not travel along
the tipping level and shaft haulage level. (parallel) to it for any great length, say, no more than 15 m at a
The analysis is undertaken over a range of infrastructure and time. Thirdly, the preferred area for the shaft collar is in a region
haulage costs, decline gradients and life-of-mine schedules. south of the main entrance road.
After a preliminary analysis of the model, it was identified that
PROBLEM DATA the faults and surface infrastructure constraints could be relaxed.
For the remainder of this paper we assume that the only the
The nominal data used for the purposes of this project is orebody standoff constraint is enforced.
summarised here and includes mine costs, decline gradients,
access points, life-of-mine schedules and no-go zones. Mine
costs are listed in Table 1. A range of values have been provided PROBLEM FORMULATION AND SOLUTION
for some of the items, so that sensitivity analyses could be PROCEDURE
undertaken to examine the effects of these parameters. All other
costs are assumed to be fixed, in the sense that they are invariant Mathematical network
to the geometry of the shaft, orepass and surrounding
infrastructure. At the time this study was undertaken, the Callie Figure 3 shows: (i) a perspective view; and (ii) a plan view of the
and Wilson Drill declines were designed to about 1400 m below shaft, orepass and surrounding infrastructure.
the surface with gradients 1:8 and 1:7 respectively. It is assumed Under this arrangement, ore from the lower levels is trucked
that both declines will continue downwards in the same manner up both declines to a level, L1, from where it is transported to the
as the current design. An alternative scheme is to have the Callie base of the shaft via a horizontal tunnel. Ore from the upper
decline gradient equal to 1:7 and the WDD gradient set at 1:6. levels is trucked down both declines to another level, L3, from
which it is transported via a horizontal tunnel to the top of the
TABLE 1 orepass. It is dropped down the orepass to L1 and transported to
Mine costs.
the shaft base. Ore from levels between L1 and L3 on the
secondary decline is either trucked up to L3 or down to L1,
Component Cost whichever is closest. Ore from levels between L1 and L3 on the
Shaft development $25 000/m, $50 000/m or Callie decline also has the option of being trucked to an
$75 000/m intermediate level, L2, which is between the top and bottom of
the orepass. It is transported to a second tipping point to the
Orepass development $1210/m orepass at L2 via a horizontal tunnel, dropped down the orepass
Level and haul drive development $3265/m to L1 and transported to the shaft base.
Haulage up decline and across $0.75/(t.km) or $1.05/(t.km) Let p1, p2 be points on the Callie decline and WDD respectively
levels at L1, p3 the point on the main decline at L2 and p4,, p5 points on
Haulage down decline $0.85/(t.km) or $1.20/(t.km) the two respective declines at L3. If levels L1, L2 and L3 are
known, the five points are fixed points corresponding to access
Orepass fit-out $1 M points on the two declines. Denote the plan location of the orepass
by x and that of the shaft by p. While the orepass is allowed to be
Access points are locations where the nominal cross-cuts positioned anywhere on the northing 9250N (which runs between
intersect the declines. Each access point is designated a nominal the two decline centrelines), the hoisting shaft must avoid the
level corresponding to its approximate Reduced Level (RL). The orebody standoff no-go region. The boundary of this orebody
surface is at approximately 1400 m RL, and access points on the standoff zone is modelled by the straight line l. Clearly, the
Callie decline extend from 340 m RL to -660 m RL at 40 m optimal position of p will be on l, rather than behind it, since in the
vertical intervals (26 points), while access points on the WDD latter case, the length of the tunnel connecting the two points can
extend from 390 m RL to 70 m RL, also in 40 m vertical be reduced by moving p onto l. The system of tunnels can be
increments (nine points). Hence in total there are 35 access points. modelled as a mathematical network T with a star topology, where
Their associated declines, (X, Y and Z) coordinates and nominal x is the centre of the star, p1,…, p5 are fixed and p is free to slide
levels are not provided in this paper. Three life-of-mine production along l.
schedules for the Callie underground mine were proposed. They For a given LOM schedule, the quantities t1,…, t5 of ore (in
are base, probable and best. Details of the three schedules are tonnes) can be computed for each fixed point p1,…, p5. For
provided in Table 2. The predicted tonnage to be accessed from example, the quantity of ore assigned to p1 is the sum of
each of the 35 access points can be determined for the three tonnages from the levels on the Callie decline below L1 and the
schedules. To avoid disruption to the shaft over the life of the levels immediately above L1 on the same decline.
mine, it must avoid impinging on several no-go regions. Firstly, Let d denote the cost per unit length of developing a tunnel,
the shaft must not be too close to the orebody. The no-go region is and h the cost per unit length of hauling a unit quantity of ore
modelled as a barrier around the Wilson and Federation orebodies. along a level tunnel. Then the sum of development and haulage
Polygons digitised around the boundaries of the orebodies were costs for the tunnel from p1,…, p5, to x is (d+hti)li, where li is the
expanded 200 m in any direction. Secondly, the shaft is required length of the tunnel from pi to x.
STRATEGIC OPTIMISATION OF A VERTICAL HOISTING SHAFT IN THE CALLIE UNDERGROUND MINE
The Fermat-Weber problem

The problem of minimising the development and haulage costs
associated with the horizontal tunnels connecting the declines,
orepass and shaft can be solved by projecting everything onto a
horizontal plane (Brimberg, Juel and Schobel, 2002). Then the
problem reduces to positioning x and p so as to minimise the
function:
6
f ( x , p ) = ∑ wi li
i=1
This is an extension of a well known mathematical problem

called the Fermat-Weber problem. The problem asks for a point,
called a Fermat-Weber point, minimising the sum of weighted
distances to a set of given points in space. In this case the orepass
position is the Fermat-Weber point, while the points p1,…, p5 and
p are the fixed points.
An added complexity is that the shaft is not fixed, but is free to
slide along the line l. We can immediately state the following
result. If x=x0 and p=pk minimise the cost function f(x,p), then
the line segment between x0 and pk is perpendicular to l. Thus,
the position of p is a function of the position of x. This can be
seen by noticing that if x0p is not perpendicular to l, then the
length of x0p can be reduced by moving p along l until it is.
Hence, for given L1, L2 and L3, the optimum network of tunnels
interconnecting the declines, orepass and shaft can be obtained
by solving an extension of the Fermat-Weber problem. This
problem can be solved iteratively using a well known descent
algorithm due to Weiszfeld (1937). For this study, the algorithm
was generalised to account for the fact that the shaft can lie
anywhere on the orebody standoff boundary and is not confined
to a fixed point. We do not give details of this amended algorithm
in this paper.
Iterative solution procedure

Initially, we proposed a simplified procedure for determining the
positions, lengths and orientations of the shaft, orepass and main
haulage drive. By this procedure, the problem was broken down
into two subproblems:
1. The optimum shaft depth was computed based on ‘vertical’
costs – decline haulage, shaft development and orepass
development – using a further extension of the
FIG 3 - (i) Perspective view and (ii) plan view of the shaft, orepass
Fermat-Weber problem, called the gradient-constrained
and surrounding infrastructure.
Fermat-Weber problem (see Brazil, Rubinstein and Volz,
2005).
Thus the associated weight is wi=d+hti. Similarly, the weight 2. A minimum-cost network interconnecting the declines,
associated with the tunnel from x to p is: orepass and shaft was constructed at the level determined
5 by the first subproblem. The optimum network was
w6 = d + h∑ ti , computed based on ‘horizontal’ costs – haul drive
i=1 development and haulage – using the Fermat-Weber
problem described above.
since that tunnel routes all the ore from the orepass base to the
shaft base. Although the simplified procedure seems to obtain good
The total cost of the mine (ignoring fixed costs) has the follow- intuitive solutions, it does not guarantee an optimal solution for a
ing components: given problem. To see this, suppose the haul level is placed at
100 m RL based on vertical costs. Numerical tests have shown
• shaft development, which depends on the shaft depth that moving the haul level to 140 m RL can reduce the cost of the
(determined by L1); network of tunnels. If the reduction in horizontal costs outweighs
• decline haulage, which depends on the levels L1, L2 and L3; the increase in vertical costs, then the solution determined by the
and simplified procedure is not optimal. Hence, we propose a
rigorous procedure which guarantees an optimal solution for a
• surrounding infrastructure development and haulage, which given problem. Refer to Figure 3. Placing the haulage drive L1 at
depend on the levels L1, L2 and L3 and the plan positions of each of the 35 levels (on both declines) from -660 m RL up to
the orepass and shaft. 390 m RL, the vertical costs can be computed. If an orepass is
Hence, the total cost of the mine can be minimised by used, it is tested at every level above the haulage drive, and if an
determining an optimal layout of the shaft, orepass and surround- intermediate tipping point is added to the orepass, it is also tested
ing infrastructure. at every level between the haulage drive and the top of the
orepass. For every possible arrangement of L1, L2 and L3, a Orepass

minimum-cost network interconnecting the declines, orepass and
shaft is computed using the amended Fermat-Weber problem.
The following effects of including an orepass in the model were
Once the cost for every mine layout has been computed, the
observed:
arrangement giving the lowest cost is selected as the globally
optimal solution. • on average, including an orepass with one tipping point
causes the optimum haul level L1 to move down by three
levels;
RESULTS
• on average, including an orepass with two tipping points
The mathematical model and algorithm described in the previous causes the optimum haul level to move down by seven levels;
section were implemented into a computer program for • including an orepass provides additional savings by reducing
determining the optimum shaft, orepass and surrounding infra- decline haulage; and
structure layout for a given set of infrastructure and haulage costs,
decline gradients and life-of-mine schedules. Thirty-six tests • provision of a second tipping point introduces further
were undertaken for each of three cases – no orepass, one savings.
orepass and one orepass with two tipping points (108 tests in Providing an orepass causes the optimum haul level to become
total). lower. This is because decline haulage is fixed for levels serviced
The gradients and cost parameters were varied across the tests. by the orepass, ie the cost of decline haulage for these levels is
Optimum haul levels and costs for the three orepass configurations independent of the shaft depth and is ignored in the model.
are compared in Table 3. Hence the shaft tipping level must be lowered to account for the
unbalanced haulage, shaft and orepass components. Including
ANALYSIS additional tipping points causes the shaft depth to increase
further. The more ore is dropped down the orepass, the deeper
In this section we analyse the results of Table 3. the shaft will be.
Shaft depth and location Gradients

The optimum haul level ranges from 390 m RL to -340 m RL The effect of increasing the Callie and WDD gradients from 1:7
(1010 m to 1740 m below the surface). A breakdown of and 1:6 to 1:8 and 1:7 respectively is to increase decline haulage
minimum and maximum haul levels is shown in Table 4. The costs. If decline haulage costs were very large compared to shaft
shaft optimally lies on the orebody standoff boundary between development, the orepass and shaft tipping levels would tend to
points (60097, 9043) and (60667, 8900), over 588 m. space themselves out over the depth of the mine, so as to
minimise the average haulage distance. Consequently the shaft
Orepass versus no orepass depth would increase. In this study, shaft development is very
If an orepass is justified, it is always economical for it to have large compared to decline haulage, and the effect of increasing
two tipping points. For the ‘base’ schedule, an orepass is not gradients has little or no effect.
justified, except if the shaft development is $25 000/m, in which The following observations were noted when the model was
case it is generally more economical to include an orepass with run with decline gradients set to zero. For the no orepass case,
two tipping points. For the ‘probable’ schedule, an orepass with the optimum haul level is 180 m RL. If an orepass with one
two tipping points is justified in all cases except PR08 and PR09. tipping point is used, the optimum haul level is -460 m RL, the
In these latter cases an orepass is not justified. For the ‘best’ top of the orepass is 180 m RL, upwards haulage is from the
schedule, an orepass with two tipping points is always justified. level at -180 m RL. If an orepass with two tipping points is used,
The orepass optimally lies on 9250N between 60250E and the optimum haul level is -160 m RL, the top of the orepass is
60755E, over 505 m. 180 m RL (haul up from -20 m RL), and the second tipping point
is at -180 m RL (haul up from -300 RL). It can be verified that,
in all three cases, the haul and orepass levels have positioned
Effects of parameters
themselves so as to minimise the total haulage cost.
We now discuss the effects of changing various parameters on
the output of the model.
Haulage costs
On average, increasing the decline haulage costs from $0.75/t.km
Shaft development and $0.85/t.km for upwards and downwards haulage to
$1.05/t.km and $1.20/t.km respectively has the effect of lowering
A breakdown of costs, averaged over all 108 tests, was
the optimum haul level by one 40 m level. The reasons are
computed. This breakdown showed that shaft development is by
similar to the reasons for the gradient effects.
far the most significant cost component, followed by decline
haulage. This explains why varying the shaft development from
$25 000/m to $50 000/m to $75 000/m has such a significant One versus two tipping points
impact on the shaft depth and the total cost of the mine. On
average, increasing the shaft development from $25 000/m to The provision of more than two tipping points could potentially
$50 000/m causes the optimum haul level to move up by three result in further savings, although the additional cost of
40 m levels. On average, increasing the shaft development from developing tunnels from the declines to the orepass must be
$50 000/m to $75 000/m causes the optimum haul level to move considered. Moreover, experiments have indicated that having
up by one 40 m level. Thus, increasing the shaft development two orepasses, one for each decline, reduces haulage and
causes a significant decrease in the shaft depth, because the development costs associated with the haulage drive which, on
resulting increase in decline haulage is outweighed by the average, account for about 12 per cent of the total variable cost
decrease in shaft development. of the mine.
STRATEGIC OPTIMISATION OF A VERTICAL HOISTING SHAFT IN THE CALLIE UNDERGROUND MINE
TABLE 3
Summary of results.
Case Schedule Shaft Haulage Haulage Callie WDD No orepass One orepass Orepass with Best option
ID development up down decline gradient two tipping
gradient points
Haul Cost Haul Cost Haul Cost
level ($M) level ($M) level ($M)
BA01 Base 25 000 0.75 0.85 1:8 1:7 158 53.4 19 53 -21 52.6 Two tipping points
BA02 Base 25 000 0.75 0.85 1:7 1:6 158 51.4 19 51.8 -21 51.6 No orepass
BA03 Base 25 000 1.05 1.2 1:8 1:7 99 60.8 -21 58.8 -101 57.7 Two tipping points
BA04 Base 25 000 1.05 1.2 1:7 1:6 99 58.3 -21 57.3 -21 56.5 Two tipping points
BA05 Base 50 000 0.75 0.85 1:8 1:7 299 83.1 219 85.3 219 85.2 No orepass
BA11 Base 75 000 1.05 1.2 1:8 1:7 299 121 278 123.7 259 123.7 No orepass
BA12 Base 75 000 1.05 1.2 1:7 1:6 398 116.9 278 120 259 120.2 No orepass
PR01 Probable 25 000 0.75 0.85 1:8 1:7 59 68.6 -21 64.6 -101 63 Two tipping points
PR02 Probable 25 000 0.75 0.85 1:7 1:6 99 65.4 -21 62.6 -101 61.5 Two tipping points
PR05 Probable 50 000 0.75 0.85 1:8 1:7 158 100.9 19 99.8 -21 99 Two tipping points
PR06 Probable 50 000 0.75 0.85 1:7 1:6 158 97.3 99 97.3 19 97.1 Two tipping points
PR09 Probable 75 000 0.75 0.85 1:8 1:7 219 131.1 99 132.7 99 132.6 No orepass
PR10 Probable 75 000 0.75 0.85 1:7 1:6 219 126.9 179 129.5 158 129.4 No orepass
PR11 Probable 75 000 1.05 1.2 1:8 1:7 158 146.5 59 145.1 -21 144.3 Two tipping points
PR12 Probable 75 000 1.05 1.2 1:7 1:6 158 141.4 99 141.3 99 141.2 Two tipping points
BE01 Best 25 000 0.75 0.85 1:8 1:7 -21 100.6 -221 89.2 -261 85.3 Two tipping points
BE03 Best 25 000 1.05 1.2 1:8 1:7 -61 125.6 -221 107.1 -341 101 Two tipping points
BE05 Best 50 000 0.75 0.85 1:8 1:7 19 135.7 -141 128.3 -141 126.5 Two tipping points
BE09 Best 75 000 0.75 0.85 1:8 1:7 77 169.5 -101 166 -61 164.7 Two tipping points
Confinement of orepass to fixed northing TABLE 4

Optimum haul levels.
Confining the orepass to be fixed on the northing between the
two declines has the effect of increasing the main haulage drive Schedule Minimum haul Maximum haul Range
development and haulage costs. The increase, however, is not level level
significant.
Base -60 m RL 390 m RL 450 m
CONCLUSIONS Probable -220 m RL 350 m RL 570 m
Best -340 m 310 m RL 650 m
In this work, we have developed and a network model for the
Callie underground mine. An algorithm was developed to
mathematically determine an optimum location and depth of a algorithm was implemented into a software product for solving
vertical hoisting shaft in the Callie underground mine. The the problem iteratively. Results were analysed over a range of
infrastructure and haulage costs, decline gradients and life-of REFERENCES

-mine schedules.
Brazil, M, Rubinstein, J H and Volz, M, 2005. The gradient-constrained
Fermat-Weber problem for underground mine design, in Proceedings
ACKNOWLEDGEMENTS 18th National ASOR Conference and Eleventh Australian
Optimisation Day (eds: L Caccetta and V Rehbock), pp 16-23.
We thank Newmont Australia Limited and the Australian
Brimberg, J, Juel, H and Schobel, A, 2002. Linear facility location in
Research Council for their joint sponsorship of this research via three dimensions – Models and solution methods, Operations
an ARC Linkage Grant. In particular, we thank Newmont for Research, 50(6):1050-1057.
suggesting and formulating the Callie shaft location study. A Kupitz, Y S and Martini, H, 1997. Geometric aspects of the generalised
simplified version of study was originally proposed by Steven Fermat-Torricelli problem, in Intuitive Geometry (eds: I Barany and
Harvey and the detailed investigation was proposed and K Boroczky), 6:55-127 (Janos Bolyai Mathematical Society).
commissioned by Andrew Fox and Ian Suckling. I thank Robert Volz, M G, 2008. Gradient-constrained flow-dependent networks for
Parr and Nadine Wetzel for their time and effort collecting and underground mine design, PhD thesis (unpublished), University of
preparing data and for their constructive feedback. Further Melbourne, Melbourne.
details of this work can be found in the author’s PhD thesis Weiszfeld, E, 1937. Sur le point pour lequel la somme des distances de n
(Volz, 2008). points donnes est minimum, Tohoku Mathematical Journal,
43:355-386.
HOME
Optimising Value for an Underground Project Configuration

T Elkington1, R Durham2 and P Myers3
ABSTRACT creating a circular problem. Lane suggested that a declining

cut-off grade strategy often delivers a high NPV. Lane (1964;
Strategic mine planning practices for open pit mines have been well
established over many years through processes generally referred to as
1988) also found that mining costs should be incorporated in the
‘optimisation’. Such methods for underground mines are not as cut-off decision for underground mines, but not open pit mines.
advanced, due in part to the added complexity and flexibility of these As such, the cut-off grade for underground mines is generally
operations. This paper presents an approach for maximising the value of higher.
a specified underground project configuration – the case study of a Nilsson (1982) proposed that optimising cut-off grade is more
hypothetical vertical gold deposit demonstrates the substantial value important for underground mines than open pit mines, because
increases and risk reduction that can be attained by optimising cut-off material left unmined underground may not be recoverable at a
grade, rather than staying with the traditional break-even approach. The later date. Additionally, underground mines will often have a
ability to optimise cut-off grade spatially is analysed and shown to higher proportion of fixed costs. Since the allocation of fixed
increase net present value by 24 per cent and decrease mine life by six costs is a contentious issue, the correct cut-off calculation for
years. The ability to rapidly optimise a single configuration enables underground mines is complex. Border (1991) recognised that
many configurations to be tested to determine the optimal project
mining cost often varies spatially in underground mines and
strategy with which to progress to a more detailed study phase.
Opportunities for optimisation and value enhancement analysed in this
should be reflected in a spatially varied cut-off grade. Hall and
paper include production rate and price sensitivity. de Vries (2003) suggested that cut-off grade could vary by
location for underground mines, even when all of the parameters
are held constant. Horsley (2005) demonstrated that low-grade
INTRODUCTION lodes can often be extracted profitably at a lower cut-off grade if
Strategic mine planning is the process of finding the there is a nearby higher grade lode that can carry the fixed and
configuration that will optimise project objectives. These development costs. Additionally, varying cut-off grade by
objectives may be value based, risk based or based on some other location can help to improve equipment utilisation and smooth
goal, depending on the company and the situation. A value based production schedules over the long-term.
objective, specifically net present value (NPV), is most often Heuristic procedures for the automation of stope design were
applied. The configuration of an underground operation consists proposed by Alford (1995) and Thomas and Earl (1999). These
of a number of key components that can be generally categorised approaches are based on ‘floating’ a minimum sized mining
into geology, mining, processing and sales. The geology shape over the resource to determine, at a given cut-off grade,
component describes the orebody, while the mining component which material should be included in the optimal mining outline.
describes the mining method, haulage plan, development concept These tools do not provide a stope design, but can be used as a
and mining rate. The processing component includes the method guide to indicate the portion of the orebody, inclusive of planned
of processing and the processing rate, while the sales component dilution, to include in the stope design. Ovanic and Young (1995)
considers the nature of the product and how it is sold. These used separable programming to optimise the placement of stopes
components each have a series of associated parameters, such as within a row of stope rings. Grieco (2003, 2007) used a mixed
the orebody geometry and its grade/tonnage relationship, costs, integer programming (MIP) formulation, incorporating grade
recoveries and revenues. uncertainty, to optimise stope design for a given risk profile.
The analysis of a project configuration requires the Brazil et al (2007) described a tool for optimising the overall
development of a strategic mine design and plan that matches the development network. This approach included the incorporation
configuration. The mine design and plan is generally of Steiner vertices to the network to minimise a combined
characterised by cut-off grade selection, stope design, haulage and development cost objective function, given a series
development design, production scheduling, and development of access points with associated tonnages. The authors also
scheduling. Each plan will have associated cash flow and value developed a tool for optimising the path of the decline through
measures, which can be used to compare alternatives. A the optimised network with a dynamic programming approach
consistent basis of comparison between project configurations is based on the optimised network. Kuchta, Newman and
essential for informed decision support. Topal (2003) used a MIP formulation to improve medium-term
The concept of varying grade to maximise discounted value scheduling at the Kiruna sublevel caving operation. The schedule
was first proposed by Berry (1922); citing the advantages of was production based, with the objective being a combination of
mining the high grade portion of an orebody first. Raising the minimising costs and deviations from desired targets. This was
cut-off grade above the zero profit point was mathematically done by optimising the timing and placement of LHDs within the
quantified by Lane (1964). Lane stated that the optimal cut-off length of the mine. A MIP formulation for sublevel stoping was
grade is influenced by the present value of the mine. This proposed by Nehring and Topal (2007). The authors incorporated
implied that the cut-off decision is dependent on the future mine activities other than production in the optimisation, such as stope
plan which was, in turn, dependent on the cut-off decision; preparation, void and filling tasks, as well as constraining the
mine against multiple fill exposures.
1. MAusIMM, Lecturer, School of Civil and Resource Engineering, Cut-off grade optimisation and schedule optimisation for a large
The University of Western Australia, 35 Stirling Highway, Crawley sublevel open stoping mine was undertaken by Smith and
WA 6009. Email: [email protected] O’Rourke (2005). They applied a mixed integer (MIP) formulation
2. MAusIMM, Associate Professor, School of Civil and Resource that simultaneously designed and scheduled stopes for a large
Engineering, The University of Western Australia, 35 Stirling orebody. The resulting solution was found to have a NPV 29.5 per
Highway, Crawley WA 6009. Email: [email protected] cent higher than that found using conventional methods, a
3. MAusIMM, Group General Manager – Engineering, Snowden substantial improvement. They also found that the life-of-mine
Mining Industry Consultants, 87 Colin Street, West Perth WA 6005. optimisation was able to achieve the target production rate for
Email: [email protected] longer than if planned with a rules-based scheduler.
T ELKINGTON, R DURHAM and P MYERS
Typically, underground strategic mine plans are completed model can provide important information to support strategic
with the sequential determination of each aspect in isolation. decision making prior to proceeding to detailed design.
Because this approach does not accurately model many The basis of this approach is that many of the aspects of the
important relationships between aspects, the resultant mine plan strategic mine plan are dependent on each other. For example,
might be a feasible solution, but is unlikely to find the theoretical stope design is dependent on the resource geometry and ore
‘optimal’ solution. Without a method to find the optimal solution, distribution, the mining method, the geotechnical constraints and
the level of suboptimality cannot be quantified. Additionally, the the cut-off grade. Additionally, development design is dependent
time- consuming and manually intensive nature of this process on the stope design and development configuration and
reduces the ability to complete multiple iterations efficiently. geotechnical constraints. This method is based upon an assumed
Currently, there is no generalised tool or algorithm available to deterministic block model and set of economic and technical
optimise each aspect of the mine configuration simultaneously. parameters. Future work to identify methods to incorporate
Optimising multiple aspects simultaneously results in an increase additional information regarding uncertainty would likely lead to
in the number of possible solutions. Carter, Lee and Baarsma a more robust strategy under a range of conditions.
(2004) suggested:
the inherent complexity of the underground Stope design related to cut-off grade
problem, will in all likelihood defy the The first step in preparing the MIP model is to determine the
development of a single solution method stoping outlines for a given mining configuration and set of
analogous to the open pit approach. geotechnical constraints at a range of cut-off values (usually
While this is the ultimate goal for strategic underground mine grade). Snowden’s Stopesizor (Myers et al, 2007) software is
planning, this paper focuses on an approach to simultaneously used for this purpose. The heuristic stope optimisation algorithm
incorporate information from multiple aspects of the mine plan involves floating a Selective Mining Block (SMB) over the
into a MIP formulation. The resultant optimisation may not find orebody. The algorithm searches for the SMB that achieves the
the theoretical ‘optimal’ solution, but supports informed decision maximum average value for the given resource. This SMB is
making. A hypothetical case study is used to demonstrate the then ‘flagged’ to be mined. The algorithm then searches for the
application of the approach. SMB that achieves the next highest average value (omitting
already flagged blocks). This process continues until there are no
more SMBs that can be added above a given minimum cut-off
MIXED INTEGER PROGRAMMING value. Flagged blocks are allocated to cut-off ‘bins’ according to
Mixed Integer Programming (MIP) is a part of the Linear the order they are included in the mining outline. A consideration
Programming (LP) family. This mathematical optimisation of the algorithm is the concept of ‘partial’ SMBs. Partial SMBs
technique solves for a number of decision variables. Each of share blocks with already flagged SMBs. The algorithm does not
these variables is given a value and is incorporated within include the grade of previously flagged blocks in the average
linearly expressed constraints. The most appealing property of grade calculation for the partial. Hence, all blocks entering the
this technique is its ability to find the optimal solution and as a mining outline must incrementally meet the cut-off value. The
result, solutions are able to be compared consistently. algorithm outputs a modified block model coded with the cut-off
MIP formulations have been used widely in open pit grade at which the block is included in the mining outline
optimisation. Ramazan and Dimitrakopoulos (2007) presented (Figure 1).
the generalised MIP formulation for optimising NPV for open
pits and tested various adjustments with respect to performance.
Caccetta, Kelsey and Giannini (1998) presented a Lagrangian A
relaxation of the MIP formulation in which mining and
processing constraints are moved into objective function as ‘soft’
constraints. Caccetta and Hill (2003) outlined a branch and cut
procedure for providing more efficient convergence of a similar
formulation.
Significant work has also been completed in regards to
optimising open pits for risk, in addition to value.
Dimitrakopoulos and Ramazan (2004) outlined a formulation of
a MIP approach which included grade uncertainty. In this B
method, the objective was to minimise a cost function comprised
of applying penalty costs to the probability that a block will have
a grade less than required. A stochastic integer programming
(SIP) formulation was proposed by Ramazan and
Dimitrakopoulos (2007). This method accounted for the risk of
violating tonnes and grade constraints when considering multiple
orebody realisations. Menabde et al (2007) provided a MIP
formulation for a number of orebody realisations and showed
significant average value increases over traditional deterministic
approaches. FIG 1 - Plan section of an example (A) resource model and
(B) mining outline (coloured by cut-off grade).
METHOD
The method proposed does not promise to provide the ‘optimal’ An appealing property of the algorithm (exploited in the
mine plan for a given project configuration. There are too many approach for this paper) is that mining outlines are nested
permutations to consider. The method used for this paper according to cut-off grade, ie high cut-off outlines are nested
considers how information from all aspects of the strategic mine within low cut-off outlines. This result is analogous to the
plan for a given configuration can be combined into a parametric generation of shells used for open pits (Whittle,
mathematical MIP model. The solving of this mathematical 1991). As this method does not account for geotechnical
OPTIMISING VALUE FOR AN UNDERGROUND PROJECT CONFIGURATION
maximum stope size constraints or support pillars. As a result,

the algorithm does not produce actual stopes; rather outlines on A
which to base practical stope designs. The algorithm does not
incorporate economic and technical parameters and hence does
not give the optimum mining outline, only a series of outlines to
be considered for further evaluation based on specified cut-off
values. However, a grade-tonnage curve based upon the stope
optimisation results provides a more accurate set of data for
economic evaluation than does the raw resource model
(Figure 2). This is largely due to the inclusion of planned dilution
and the exclusion of isolated above cut-off value resource blocks.
FIG 2 - Resource – mining inventory comparison (annotated by

FIG 3 - Level development design for (A) 1 g/t cut-off and
cut-off grade).
(B) 7 g/t cut-off.
Development design related to cut-off grade Hence, the global optimisation focuses on cut-off grade and
schedule optimisation, whilst ensuring that the stope and
Development design is strongly dependent on stope design, or development designs are consistent with the cut-off grade(s)
mining outline in this case. Conceptual development designs selected in the optimisation (Figure 4). The objective of the
were created for each stope design (one stope design for each optimisation is to select the cut-off grade policy and schedule
cut-off grade). These designs were completed manually, based on that maximises NPV, ie what should be mined and when it
a given development configuration. An extension to this study should be mined. It is recognised that NPV might not be the
would be to optimise the development network/decline using primary objective of every underground operation, and that
methods such as those outlined by Brazil et al (2007). The optimisations based on other objectives such as total cash flow,
development designs were completed in order to determine the mine life, or risk could be considered.
relationship between development metres and cut-off grade on Completing a multi-period MIP optimisation on a block-
each level (Figure 3). This allows a given stope design to be by-block resolution is beyond current hardware and software
matched with a valid development design. capabilities. In order to make it practical to solve the problem, it
is necessary to create a smaller number of activities to schedule.
Strategic schedule optimisation This requires the grouping of blocks. An example of this is given
in Stone et al (2007) for an open pit application. Grouped blocks
The schedule optimisation used in this method incorporates the will carry the same (averaged) properties, and resolution is lost.
information gained through stope and development design Grouped blocks also report the same output, including their
analysis at a number of cut-off grades. It is assumed that for a ultimate destination and the time period in which this occurs.
given project configuration, development design is dependent on Because of this, it is important that resolution is managed
stope design and stope design is dependent on cut-off grade. efficiently by grouping blocks that are likely to be mined
FIG 4 - Optimisation concept.
together and sent to the same destination. The grouping process

used for the case study in this paper is to group blocks by level
and cut-off grade to form stoping activities. It is assumed that A B
material on a given level will be extracted at approximately the
same time, and that material in each cut-off grade interval will be
sent to the same destination. Grouping is also applied for
development activities. Decline development and level access
development (access crosscut and footwall drive) is grouped by
level. Finally, variable development (development driven by
tonnage/cut-off grade) is incorporated within the stoping
activities.
For the MIP model, a decision variable is created for each
stoping and development group in each period. Each decision
variable measures the proportion of the activity that is completed
in the given period. NPV theory is based upon the concept that
cash flow received today is more valuable than an equivalent
cash flow received in the future. As the objective of the MIP
optimisation is to maximise NPV, the objective value associated
with each activity is equal to the cash flow (positive or negative)
discounted at a specified rate. The cash flow associated with each
activity is calculated from the economic and technical
parameters associated with the activity. A range of important
constraints are applied in order to reflect the mining practicality
of the given project configuration, as well as to control the
optimisation. FIG 5 - View of orebody looking south coloured by (A) grade,
The constraints that reflect the project configuration include a (B) mining front.
minimum and maximum limit on mining/processing tonnages and
development metres per time period. Note that for a configuration
with a single processing plant, the mining and processing limits in thickness with a minimum separation of 40 m between the
would usually be the same or similar. Development constraints can lodes. The rock conditions, vertical extent and close proximity of
be grouped according to vertical and horizontal development to the lodes suggest that open stoping would be an appropriate
reflect the fact that different equipment is sometimes used for each mining method. Mining inventories incorporating planned dilution
purpose. These constraints can be applied globally or locally (to a were generated for each cut-off grade with Snowden’s Stopesizor
subset of activities). For example, while the overall mining limit software tool. Cut-off grades considered in this analysis were 1 g/t
might be 1 Mt/a, the maximum that can be extracted from a single to 10 g/t in 0.5 g/t increments. The minimum mining shape was
mining area might only be 500 kt/a. Additionally, blending assumed to be 20 mE × 20 mN × 20 mRL, based on preliminary
constraints and selling constraints can be applied to enforce head geotechnical analysis. Applying this shape to the orebody
grade bounds and total metal output. Wooller (2001) provides an resulted in significant planned dilution and the divergence of the
interesting implementation of selling constraints to model mining inventory from the resource (Figure 2). This result
processing plant performance. reflects the bulk nature of the mining method.
The project configuration will also outline the direction of A basic development design was completed on each level for
mining. This is imposed as a series of sequencing constraints each cut-off grade (Figure 3). The decline was placed on the
based on precedence relationships between activities. These western side of the orebody and the ore was accessed via a drive
relationships can link development and stoping activities, located centrally to the two lodes. The designs were used to
development activities only or stoping activities only. While derive average decline metres and access development for each
geotechnical constraints are not directly applied in this method, level, as well as the relationship between the tonnes in the
sequencing constraints can be used to simulate this. Integer mining inventory and the development required on each level.
decision variables are introduced to monitor whether all activities Decline and level access development was assumed to be
precedent to a given activity have been completed.
independent of cut-off grade.
The optimisation is also controlled by two other important
The underground optimisation was based upon optimising
types of constraint. Firstly, constraints are imposed to ensure
each activity is only completed once. Secondly, constraints are value for a 1 Mt/a production rate. The overall mine extraction
used to allow only one cut-off grade to be selected for a given sequence is top-down via a bottom-up stoping method from four
activity. Integer decision variables are used to manage these production fronts, each capable of providing 500 ktpa.
constraints – the solution to the optimisation problem can be Consequently, at least two fronts need to be operational in any
interpreted into a mining and development schedule. period to achieve full production. The top level in each
Consequently, each block can be coded with the time period in production front remains as a pillar, which can be extracted once
which it is mined. This enables the output to be exported to mine all levels above it are completed. A maximum decline
design or scheduling packages for further detailed analysis. development rate of 1500 M/a and horizontal development rate
of 5500 M/a was assumed. Development was constrained so that
the decline and level access must precede any level development
CASE STUDY on a given level.
The case study project used to demonstrate the application of the An initial optimisation was conducted without any assumed
approach is a hypothetical gold project. production fronts or stoping activity sequencing. This allowed
the optimisation to reveal the areas where the highest values
Background exist. These areas indicate the best locations to start production
from, if geotechnical constraints were not important. This
The orebody is a large dual lode gold deposit (Figure 5a). The two represents an upper bound solution to guide future strategy.
lodes are near vertical and cover 900 m in vertical extent. The Production fronts were allocated based on the concept of trying
orebody lies beneath 60 m of overburden. The lodes average 40 m to replicate the upper bound solution with practical mining
constraints. This was done by starting production for a

production front on a level that is extracted early in the upper
bound solution. Likewise, crown pillars for each production front
were located at levels taken late in the best case solution.
Production front allocation is shown in Figure 5b.
Results
This project configuration was analysed under two different
assumptions regarding cut-off grade policy. Firstly, the
configuration was optimised for each cut-off grade, assuming that
the same cut-off grade is applied on every level. Secondly, the
configuration was optimised whilst allowing cut-off grade to be
varied between levels. An important assumption made for this
optimisation was that the stope design on a level is independent of
the design of all others. This allowed for the case of varying
cut-off grade by level to be analysed. Additionally, the
conservative assumption was made that once mining is completed
FIG 7 - Single cut-off grade 5 g/t production schedule.
on a level to the selected cut-off grade, the material beneath this
cut-off grade is sterilised.
Single cut-off grade

A single cut-off is perceived to be the simplest approach in terms
of design and operation. Figure 6 shows that NPV is maximised
at a cut-off grade of 5 g/t with 19.3 Mt of ore.
FIG 8 - Cumulative discounted cash flow schedule comparison.
significant. Though this will result in a shorter mine life, the

reduction is taken from periods that do not provide meaningful
contributions to discounted value. It is expected that the shorter
FIG 6 - NPV curve. the mine life, the closer the optimal single cut-off grade will be
to the break-even cut-off grade.
The break-even cut-off, most commonly applied in operating
practice, is 3.5 g/t. At the break-even cut-off grade there is a Spatially varying cut-off grade
17 per cent drop in NPV accompanying a 4.6 Mt increase in ore
processed. In this case, the break-even cut-off strategy promotes Testing Horsley’s (2005) findings that value can be improved by
mining longer for less NPV. The schedule, at a 5 g/t cut-off, is varying cut-off grade spatially, the optimisation was allowed to
shown in Figure 7. One of the key results of the schedule allocate any of the 19 cut-off options on each level – this
optimisation is that production front four does not commence optimisation resulted in an NPV improvement of seven per cent
until year 11 of the operation. The decline is capable of reaching (Figure 6) over the single 5 g/t cut-off grade. The tonnage
this front by year six. The optimisation was able to resolve the
associated with this option is 17.4 Mt. The seven per cent
compromise between grade differentials and the added cost of
development in deciding to delay this development. An increase in value is gained through mining at a higher average
additional characteristic of the optimisation was the ability to grade in periods 4 - 11 (Figure 9). Periods 1 - 3 are constrained
maintain a sustainable production rate until the end of the mine by the development to areas that can be mined and hence average
life. This was due to the multi-period optimisation being able to grade is low. The higher average grade, due to increased cut-offs
identify and rectify bottlenecks by adapting the strategy in early in the periods 4 - 11, results in greater cash flows when
periods. discounting does not impact so greatly. At the end of the mine
The source of the increased value associated with an increase life, additional material is able to be mined at a lower cut-off
in cut-off grade above break-even is opportunity cost. grade than 5g/t in order to maximise extraction, due to the lesser
Discounting makes cash inflows in later periods (of a long life impact of discounting.
operation) worth little in today’s terms. Figure 8 shows the Overall, this solution represents a 24 per cent increase in value
minimal value contribution of the final five years of the over the break-even solution with a mine life of six years less. This
break-even option. Hence, it is often (and demonstrated in this is assuming that the break-even schedule has been optimised for
case) appropriate to raise the cut-off grade to achieve a higher NPV, which it rarely is. This method could potentially find a
average grade in those periods when discounting is not as greater NPV improvement over a non-optimised break-even plan.
operation being development-constrained at the time this area is

to be mined.
MULTIPLE CONFIGURATION ANALYSIS

The concept of rapid optimisation of value for a single project
configuration allows a number of configurations to be
subsequently optimised and compared in an efficient manner. The
following analyses describe how the optimisation methodology
can be extended to allow for greater information in strategic
decision making.
Production rate optimisation

Production rate has a number of impacts for a new underground
project. An increase in production rate can result in an increase
in upfront capital cost due to plant and mining fleet size, a
FIG 9 - Spatially varying cut-off grade production schedule.
decrease in unit mining and processing costs due to economies of
scale, an increase in fixed costs associated with maintaining
Lane (1964; 1988) suggested that cut-off was time-dependent larger fleet and plant, an increased presence of bottlenecks due to
rather than spatially dependent. Figure 10 shows that levels that additional equipment, an increased effort to maintain a
were mined early had a much higher cut-off grade (up to 7.5 g/t) sustainable production rate and a shortened mine life. Figure 12
than those level mined later (down to 3.5 g/t), due to the effect of shows a contour plot interpolating NPV for each combination of
discounting. The only exception is the first few years, which 19 cut-off grades (single cut-off) and eight production rates
have lower cut-offs due to the lack of available ore. The general between 0.8 Mt/a and 1.5 Mt/a. The optimal NPV is achieved at
trend for each production front was for the cut-off grade to
a cut-off grade of 5.5 g/t and production rate of 1.3 Mtpa
decrease as mining proceeds. In fact, plotting the start time
(marked with a ‘×’ in Figure 12). The contour line immediately
against cut-off grade for each level resulted in a clear negative
linear relationship (Figure 11). This indicates that in this case, around that point represents the solutions within five per cent of
time is a more important driver of cut-off grades than location. the optimal solution and indicates a number of acceptable
Production front one has much lower cut-off grades due to the solutions in this case. The break-even cut-off strategy of 3.5 g/t
does not lie within this optimal region. In fact, depending on the
production rate, this strategy will be somewhere between ten per
cent and 35 per cent suboptimal.
FIG 10 - Spatially varying cut-off grade – optimal solution.

FIG 12 - Interpolated relative NPV.
Strategy sensitivity to price

The most relevant question in terms of price sensitivity is:
which strategy, selected today, allows the project
to be optimal or near optimal under a range of
commodity price scenarios?
The ability to quickly optimise for a single project configuration
enables a number of commodity price scenarios to be considered.
Scenarios of price between -30 per cent and +30 per cent (relative
to the base case price) were considered for each of the 19 cut-off
grade options (single cut-off). The results (Figure 13) are plotted
in a contour graph, with an interpolation of NPV relative to the
maximum NPV for each price scenario. An important result of this
analysis is that the selection of a cut-off grade of 5.5 - 6 g/t would
be, at most, five per cent suboptimal under all of these price
FIG 11 - Relationship between start time and cut-off grade (one conditions. This cut-off grade would be considered quite resilient
sample from each level). to price uncertainty. The modest negative relationship between
Border, S, 1991. Optimisation of cut-off grades during design of

underground mines, in Proceedings Mining Industry Optimisation
Conference, pp 21-27 (The Australasian Institute of Mining and
Brazil, M, Lee, D, Rubinstein, J, Thomas, D, Weng, J and Wormald, N,
2007. Optimisation in the design of underground mine access, in
R Dimitrakopoulos), pp 141-144 (The Australasian Institute of
Caccetta, L and Hill, S, 2003. An application of branch and cut to open
pit mine scheduling, Journal of Global Optimisation, 27:349-365.
Caccetta, L, Kelsey, P and Giannini, L, 1998. Open pit mine production
scheduling, in Proceedings Third Regional APCOM Symposium –
Computer Applications in the Minerals Industries International
Symposium, pp 65-72 (The Australasian Institute of Mining and
FIG 13 - Interpolated NPV (relative to the maximum NPV at
Carter, P, Lee, D and Baarsma, H, 2004. Optimisation methods for the
each price).
selection of an underground mining method, in Proceedings Orebody
Modelling and Strategic Mine Planning, pp 7-12 (The Australasian
price and optimal cut-off grade indicates that for this case, an Institute of Mining and Metallurgy: Melbourne).
optimised mine strategy is quite resilient to changes in price. The Dimitrakopoulos, R and Ramazan, S, 2004. Uncertainty-based production
break-even cut-off grade strategy of 3.5 g/t for the base case is scheduling in open pit mining, Transactions of the SME, pp 106-112.
between 11 per cent and 100 per cent suboptimal under the Grieco, N, 2003. Risk analysis of optimal stope design: Incorporating
various price conditions. A break-even strategy not only misses uncertainty, Master of Philosophy thesis, The University of
value opportunity, but is also a higher risk option. Queensland.
Grieco, N and Dimitrakopoulos, R, 2007. Grade uncertainty in stope design:
CONCLUSIONS Improving the optimisation process, in Orebody Modelling and Strategic
Mine Planning, second edition (ed: R Dimitrakopoulos), pp 167-174
This paper presents an approach to incorporating information from (The Australasian Institute of Mining and Metallurgy: Melbourne).
a number of aspects of underground strategic mine planning into a Hall, B and De Vries, J, 2003. Quantifying the economic risk of
single optimisation. The case study found that substantial gains in suboptimal mine plans and strategies, in Proceedings Mining Risk
value could be accessed by mining in a more strategically focused Management Conference, pp 59-69 (The Australasian Institute of
way. In particular, it showed that a break-even cut-off grade Mining and Metallurgy: Melbourne).
strategy encourages mining more material for less value. Gains in Horsley, T, 2005. Differential cut-off grades, in Proceedings Ninth
value of up to 24 per cent were found through optimising cut-off Underground Operators’ Conference, pp 103-109 (The Australasian
grade spatially and with time. This maximum gain was achieved Institute of Mining and Metallurgy: Melbourne).
while mining for six years less. Comparatively greater gains were Kuchta, M, Newman, A and Topal, E, 2003. Implementing a production
likely, given that traditional break-even cut-off grade schedules are schedule at lkab’s Kiruna mine, INTERFACE, 34(2)124-134.
not optimised. Lane, K, 1964. Choosing the optimum cutoff grade, Quarterly for the
The optimal mine plan for the base case was also found to be Colorado School of Mines, 59(4):811-829.
much more resilient to price uncertainty than a break-even cut-off Lane, K, 1988. The Economic Definition of Ore: Cut-Off Grades in
strategy. Under a ±30 per cent price range, the optimal base case Theory and Practice (Mining Journal Books: London).
cut-off is at most five per cent suboptimal. For the same range, the Menabde, M, Froyland, G, Stone, P and Yeates, G A, 2007. Mining
break-even strategy is between ten per cent and 100 per cent schedule optimisation for conditionally simulated orebodies, in
suboptimal. The gain in value and reduction in risk were made Orebody Modelling and Strategic Mine Planning, second edition (ed:
with little, if any, additional planning effort. This gain does not R Dimitrakopoulos), pp 379-384 (The Australasian Institute of Mining
include the potential for additional improvements to be made from and Metallurgy: Melbourne).
considering other configurations that could increase value or Myers, P, Standing, C, Collier, P and Noppe, M, 2007. Assessing
mitigate risk. Given the current economic climate, the need to underground mining potential at Ernest Henry mine using
optimise the utilisation of mineral deposits has become essential. conditional simulation and stope optimisation, in Orebody
Optimisation methods such as the one described in this paper Modelling and Strategic Mine Planning, second edition (ed: R
should be applied to all mining operations. They allow many Dimitrakopoulos), pp 191-200 (The Australasian Institute of Mining
project configurations to be compared on a consistent basis, and and Metallurgy: Melbourne).
provide confidence that the best strategy has been found. The Nehring, M and Topal, E, 2007. Production schedule optimisation in
optimal solution can then be used to drive more detailed design underground hard rock mining using mixed integer programming, in
and scheduling. Proceedings Project Evaluation 2007, pp 169-178 (The Australasian
ACKNOWLEDGEMENTS Nilsson, D, 1982. Optimum cut-off grades in underground mining,
Canadian Mining Journal, 103(10):65-70.
The authors would like to thank Snowden Mining Industry
Consultants for their support of this research and for applying a Ovanic, J and Young, D, 1995. Economic optimisation of stope geometry
similar approach for their in-house Evaluator tool. Additionally, using separable programming with special branch and bound
techniques, in Proceedings Third Canadian APCOM, Montreal,
we would like to thank R Secis and A Earl from Snowden
pp 22-25.
Mining Industry Consultants for reviewing this paper.
Ramazan, S and Dimitrakopoulos, R, 2004. Recent applications of
operations research in open pit mining, Transactions of the SME,
REFERENCES 314:73-78.
Alford, C, 1995. Optimisation in underground mine design, in Ramazan, S and Dimitrakopoulos, R, 2007. Stochastic optimisation of
Proceedings APCOM XXV, Brisbane, pp 213-218. long-term production scheduling for open pit mines with a new integer
Berry, E, 1922. Present value in its relation to ore reserves, plant capacity programming formulation, in Orebody Modelling and Strategic Mine
and grade of ore, American Institute of Mining and Metallurgical Planning, second edition (ed: R Dimitrakopoulos), pp 385-392 (The
Engineers, July, 11-16. Australasian Institute of Mining and Metallurgy: Melbourne).
Smith, M and O’Rourke, A, 2005. The connection between production Whittle, J and Rozman, L, 1991. Open pit design in the 90s, in
schedules and cut-off optimisation in underground mines, in Proceedings Mining Industry Optimisation Conference, pp 13-19
Proceedings APCOM XXXII. (The Australasian Institute of Mining and Metallurgy: Melbourne).
Stone, P, Froyland, G, Menabde, M, Law, B, Pasyar, R and Monkhouse, Wooller, R, 2001. Cut-off grades beyond the mine – Optimising mill
P, 2007. Blasor – Blended iron ore mine planning optimisation at throughput, in Mineral Resource and Ore Reserve Estimation – The
Yandi, in Orebody Modelling and Strategic Mine Planning, second AusIMM Guide to Good Practice (ed: A Edwards), pp 459-468 (The
edition (ed: R Dimitrakopoulos), pp 133-136 (The Australasian Australasian Institute of Mining and Metallurgy: Melbourne).
Thomas, G and Earl, A, 1999. The application of second-generation
optimisation tools in underground cut-off grade analysis, in
Proceedings Strategic Mine Planning, pp 175-180 (Whittle
Programming Pty Ltd: Box Hill).
HOME
Classification of Mining Methods for Deep Orebodies

V Oparin1, A Tapsiev2 and A Freidin1
ABSTRACT support method in the course of mining’. We acknowledge the

effectiveness and advantages of both classifications, though some
Classifications of mining methods date back to the 1960s and use the
following criteria: type and size of a mineral deposit, mined-out space
comments should be made. First, these classifications appeared
support, state of a working excavation, type of a face, roof support, etc. in the mid-1960s. Mining methods have been continuously
Since that time mobile mining machinery has appreciably advanced, and evolving since then, and many of them have dwindled, especially
some mining methods have lost their importance. Mobile mining for medium thick and very thick ore deposits. These days,
equipment is effective in definite mining methods: in open mined-out underground mining practice involves lively mobile machinery
space, mining with backfill and combined schemes. The transition to sets, which has drastically simplified preparatory and
deep mining inevitably results in the sharply worsened technical and development operations as well as stoped excavation, boosted
geomechanical conditions, and some well-established techniques for safe mining works and minimised hand-labour amount due to the
mining at rock burst hazardous deposits can not be considered overall mechanisation.
self-sufficient. This paper puts forward a classification of deep mining Highly laborious mining with portable equipment is now used
methods, considering rock pressure control. The classification involves in the excavation of very thin orebodies with high useful mineral
three classes: mining with backfill, mining with caving of overlying rocks
content. This method and equipment may only be applied for
and combined mining with backfill and caving.
actual mining of medium thick and very thick ore deposits if
low-paid manpower is in employment, mineral resources are of
INTRODUCTION low value and the total enterprise is of no consequence to be
re-constructed or technically modernised. The analysis of the
Many foremost Russian and foreign researchers and mining mobile machinery operation shows that they are efficient in a
engineers have been scrutinising methods of grouping existing limited number of mining methods, namely:
underground mining methods. More than 60 published • mining with backfill (longwall and chamber mining,
classifications have undergone critical analysis in fundamental including the chamber-and-pillar method);
monographs by Trushkov (1947), Agoshkov (1965), Baikonurov
(1969) and Imenitov (1978). Trushkov conceded on inexpediency
• mining without backfill (chamber-and-pillar and flat-bottom-
chamber mining);
to try and embrace all coal and ore mining methods in a uniform
classification due to ‘crockness and inconsistency of these • mining with ore and rock caving (block caving and,
methods, and their not-hands-down use in practice’ (Trushkov, especially, sublevel caving with frontal and frontal-areal ore
1947). Unlike coal formations, orebodies have multiple sizes and drawing); and
shapes, occurrence conditions, physico-mechanical properties, • combined mining (with backfill and caving).
mineral composition and mineral values, which gave rise to a
range of 170 to 200 variants of ore mining schemes applied in The gained engineering level of underground mining with
underground mines in the 20th century (Trushkov, 1947; mobile machinery and successful robotic automation of ore
Agoshkov, 1965). excavation processes at some enterprises (Bronnikov, 1982;
The totality of underground ore mining methods caused Oparin, 2007) will lead to further universalisation of mining
difficulties with scientific grouping, however, it also fostered methods. It is also important to consider that underground mining
many characteristics for classifying them. Different researchers gets deeper by 20 - 40 m annually, and now reaches nearly a four
assumed different classification criteria (please refer to km depth in South Africa and India. More than ten underground
Baikonurov, 1969); for example, the type and size of a deposit mines in USA, Canada and South Africa are below 2.5 km depth.
(Crane, G J Young, I Pokrovsky, G E Bakanov and L I Baron); Deep mining conditions are very difficult: rock temperature is
mined-out space support (F W Sperr, Y H Rayt, E C Mitke and V higher, the physico-mechanical properties of rocks change, angles
N Semevsky); working excavation support (J F Clelland, C F of rock mass movement flatten and rock pressure grows and
Sackson and E D Gardner, USA Mining Bureau’s classification, manifests itself dynamically. More and more deposits acquire the
N I Trushkov, N A Starikov, G N Popov and V R Imenitov); category of rock burst hazardous. Mining theory and practice have
working excavation state (M I Agoshkov and R P Kaplunov); elaborated measures for minimising the hazard of rock pressure
mined-out space support during mining and backfill (V T (Bronnikov, 1982) to:
Markelov); type of a face, roof support, block undercutting • prepare and develop orebodies in a minimum fractured rock
technique (American Institute of Mining Engineers and mass so that no stress concentration areas arise (pillars,
Metallurgists); direction of mining advance and backfill method overhangs) resulting in burst-like rock failure;
(Baikonurov, 1969); mining stages (Steshenko). The above-listed
criteria rank and characterise mining methods in varying degrees • orient development workings and stopes along the maximal
and yet are of little use in actual comparison and selection. In effective stresses; and
Russia, the classifications by Agoshkov (1965) and Imenitov • to advance stoped excavation as divaricated fronts, or as a
(1978) have found deserved recognition and wide application in single front towards the wing of a level.
underground mine planning. Agoshkov grouped mining methods
These measures collide with many methods of underground
by ‘the state of the working excavation in the course of actual
mining, for example, with ‘ore shrinkage’ and ‘mining without
mining’ and Imenitov used the criterion of ‘working excavation
backfill’ by Agoshkov (1965), or ‘mining with naturally
supported working excavation’ by Imenitov (1978).
1. Institute of Mining, Siberian Branch, Russian Academy of Sciences, Different methods of mining with ore shrinkage that are in wide
54 Krasny Prospect, Novosibirsk 630091, Russia. application in thin and steep orebodies, involve ascending stoping
2. Institute of Mining, Siberian Branch, Russian Academy of Sciences, towards a diminished pillar. This is not permissible for us in rock
54 Krasny Prospect, Novosibirsk 630091, Russia. burst hazardous rock masses. For instance, at Yuzhny complex
Email: [email protected] deposit (GMK Dalpolimetal Co), where thin and strong ore veins
V OPARIN, A TAPSIEV and A FREIDIN
occur in hard host rocks prone to elastic energy accumulation and • the classification is for industrial-scale mining, with only the
dynamic failure (Pilenkov, 1990), series of rock bursts with severe exclusion of thin ore deposits where portable mining
after-effects took place at a depth of 150 - 200 m from the surface equipment is unavoidable due to space-limited environment;
when mining with ore shrinkage approached the crown pillar. The
• the secondary criteria should be disregarded, for example,
ground control is extremely limited when mining is carried out
when strike-line or dip-line mining is governed not by the
without backfill, and the roof is supported with ore pillars and mining method, but by the parameters of the deposit and its
lining. In this case, the growing mined-out space loses its mining-technical and geomechanical conditions;
equilibrium, dynamic rock failure conditions originate and
large-scale rock fall becomes highly possible. A representative • the systems with support are withdrawn from the
illustration of the above is the underground mines at Zhezkazgan classification as all of the known mining methods involve the
cupriferous sandstone deposit (Yun, 1997), where the chamber- support; and
and-pillar method involves retaining enlarged inner-chamber and • the basic classification criterion allows entering any new
panel pillars (up to 40 per cent and above of a panel area). mining method to the classification.
Nonetheless, starting from a 450 m depth, falls and burst-like
Thus, in accordance with the effective ground control methods,
failure of pillars occur again and again. The calculations indicate
we divide deep mining system into three classes:
that for chamber-and-pillar mining at a depth of 1000 m, more
than half-reserves should be left as pillars for roof support safety 1. Class I – mining with backfill,
(Bronnikov, 1982; Zamesov, 1979). 2. Class II – mining with overlying (host) rock caving, and
3. Class III – combined mining with backfill and caving.
METHOD OF CLASSIFICATION
Classes I and II are commonly known from technical literature
An important point in Agoshkov (1965) and Imenitov’s (1978) and the practice of underground mining. We would only note the
mining method classifications is the lack of the systems of undesirable large-scale blasting during mining with caving and
chamber mining with backfill. These systems are of use in frontal ore drawing at rock burst hazardous deposits. Large-scale
longwalling and chamber- and-pillar working at 1000 m depth and explosions release huge energy, which initiates displacement of
deeper. The underground chamber-and-pillar mining with backfill large rock blocks, causes dynamic events in the rock mass and
is applied extensively in Canada (Inco, Geco, Kidd Creek Mines), failure of the weakened ore drawing-off level. These events are
Finland (Outokumpu) and Australia (Mount Isa) (Khomyakov, frequent at the Tashtagol iron-ore deposit, where the mining
1984), as well as in the former USSR countries (Gaisky, method includes block caving, vibration areal drawing of ore and
Leninogorsky, Zyryanovsky Mines, etc). Some engineers think large-scale block breakage. The Kiruna Mine in Sweden has the
these geo-technologies are useful in mining no deeper than 600 - same rock mechanics and identical depth of mining, but with
800 m, due to the growing bearing pressure after the first sublevel caving, slice breaking and frontal ore drawing. As a
chambers are extracted (Bronnikov, 1982; Zamesov, 1979). Failure result, dynamic pressure manifestations are absent here.
of pillars and excavations at ore drawing-off level necessitates Therefore, we include only slice breaking with frontal and
introduction of longwall chamber methods that proved themselves frontal-areal ore drawing into our classification (Freidin, 2008).
in the Norilsk mines. Thus, for rock burst hazardous ore deposits, The technologies involved with the combined ground control
we think it is possible to characterise the allowable mining methods are as follows:
systems by:
• paste filling costs are expensive, and
• the principal ground control method for the mined-out space,
• mining of low and mid-grade ores calls for cheaper methods.
• the stoping excavation advance direction, and
In particular, partial backfill is admissible in thick and very
• the ore breaking and drawing method. thick gentle and weak-inclined deposits, with the purpose of
The ground control is the governing factor for any deep mining providing the roof support and smooth sagging, limited dis-
technology. It is always present in the description and in the name placements and strains of the host rocks. In this case, the filled
of a mining method. We assume it the basic classification areas alternate with the areas of mining under the caved roof
criterion. formed as a stable dome. First, mining captures areas with
It is natural that the principal ground control method is artificial pillars, and secondly, the areas with temporarily
contributed to by the known design solutions oriented to enhance retained ore pillars. The smooth sagging of the roof is ensured
strain capacity of a rock mass and to monitor stresses in the with the help of the artificial pillars, the sides of which are
near-face area. Such supportive measures include, for example, supported by caved overlying rocks. Parameters of these pillars
overworking (undermining) of the orebody, advance boring of should limit the rate of the roof sagging to the values of mining
larger diameter holes, confined explosions, directional hydro- with full backfill. The width of the temporary ore pillars is
fracturing, etc. These measures are feasible with any mining designed with allowance for the stable span. Depending on the
method and are not the attributes of a particular recovery stability of ore and host rocks, and orebody thickness, mining
mechanism. The second-listed characteristic is often present in the with backfill may include horizontal slicing with ascending (or
name of a mining method and gives a description of it. For descending) pillar mining, and the temporary pillar mining may
example, the horizontal slicing with backfill may be ascending and involve sublevel caving (for very thick orebodies) or block
descending. Steep rock burst hazardous ore veins are mined with caving (for thick deposits). The offered classification of the
sublevel stoping to the dip or on the strike. The third characteristic mining methods is presented in Table 1. The information in this
in our list details the way of ore drawing, namely, areal (from table in terms of mining methods is further depicted in a series of
inclines), frontal or frontal-areal. Pillar mining may involve VCR, related figures.
parallel or fan-patterned boreholes. So, the second and third
criteria can and must serve for the grouping of the mining CONCLUSIONS
methods.
This classification of mining methods for deep orebodies has
We are proposing a variant of classification of the mining been offered for the first time, and is based on the underground
methods for rock burst hazardous ore deposits based on the mining practice in Russia, in particular, in Norilsk Region. The
following principles: authors think the classification may successfully be used for
• the mining methods should be obligated to meet the stringent selecting and optimising the extraction of rockburst-hazardous
standards of operations in the conditions of rock burst hazard; ore and rock masses.
CLASSIFICATION OF MINING METHODS FOR DEEP OREBODIES
TABLE 1
Classification of mining methods for rock burst hazardous ore deposits (table includes references to the corresponding figures).
Class Group Mining Application conditions Advantages Disadvantages

method
1 2 3 4 5 6
I. Ground A. Horizontal 1. Horizontal Any geometry of a deposit. Medium Simple geometry of stoping Strict regulation of slicing.
control with slicing with slicing and and greater thickness. Average faces, high extraction index. In case of high rock pressure,
backfill backfill ascending stability ores, mid-stable to unstable Mobile machinery use. Sufficient overworking is desirable,
mining host rocks. Medium to high-grade stoping productivity. Possible which is highly laborious.
ore. Unlimited depth of application. separation of dirt rocks for Organizational difficulties in
backfill. stoping automation
(Figure 1).
2. Horizontal Any geometry of a deposit. Medium The same as in ascending The same as in ascending
slicing and and greater thickness. Very unstable mining. Safe mining in heavily mining. High labour intensity
descending and heavily dislocated ores and host dislocated orebodies. and cost of backfilling
mining rocks. High-grade ore. Unlimited (Figure 2).
depth of application. Open cast and
underground mining.
3. Horizontal Gently dipping medium thick and Ascending and descending Driving and backfill of extra
slicing with very thick orebodies. Mid-stable to mining of main reserves and mine workings in the bottom
pillars on soft unstable ores and host rocks. temporary pillars. Increased of temporary pillars. Stoping
basis High-grade ore. Depth of application mining front and more stopes in stages (Figure 3).
is to be determined. under development. High
extraction index.
4. Room work Mid-thick to very thick deposits. Higher quality and completeness Mining in stages. Higher
and slicing Any dipping. Mid-stable to unstable of pillar extraction. labour intensity and lower
ores. Mid-stable host rocks. Medium stoping production (Figure 4).
to high-grade ores. Depth of
application is to be determined.
5. Longwall Horizontal and gently dipping Selective mining. Elongated Arduous working conditions
mining orebodies. Very thin to 3 - 4 m thick stoping front. High extraction and high labour intensity. Use
ores. Any stability ores, mid-stable index. of portable equipment, limited
host rocks. Paste backfill, packs, use of mobile machinery
stony and artificial blocks. (Figure 5).
High-grade ores. Unlimited depth of
application.
B. Room 6. Continuous Medium thick to very thick deposits. High stoping productivity. High ore dilution with
work mining chamber Any dipping. Medium stable ores Automation of principal cement-containing backfill.
mining and host rocks. Depth of application operations in stoping. Limited mining front
is to be determined. Medium to (Figure 6).
high-grade ores.
7. Chamber- Medium thick to very thick deposits. High stoping productivity. Mining in stages. Higher
and-pillar Any dipping. Medium stable ores Automation of principal requirements to ore extraction
mining (areal, and host rocks. Medium and operations in stoping. Extracted standards. Probable failure of
frontal and high-grade ores. Areal, frontal and ore quality control. secondary chambers (pillars)
frontal-areal frontal-areal ore drawing depending in case of high pressure. High
ore drawing) on the orebody thickness. ore dilution (Figure 7).
II. Ground A. Block 1. Block caving Very thick, high dip outcropping or High stoping productivity. Probable hanging-up of host
control with caving with slicing close to surface ores. Limited Relatively low mining costs. rocks with following
overlying and frontal application in gently dipping and High automation of stoping and burst-like destruction of large
rocks caving (frontal-areal) inclined ore deposits with weak and use of mobile machinery. structural rocks. Forced
ore drawing unstable overlying rocks. Hard Sufficient extraction index with ventilation of ore drawing-off
medium stable and stable, low and frontal-areal ore drawing. workings (Figure 8).
medium grade ores. Limited depth Ventilation due to mine
of application. Unstable and medium depression.
stable host rocks. Ores are non-liable
to spontaneous ignition and caking.
B. Sublevel 2. Sublevel Thick and very thick, high dip Separation of dirt bands to Probable hanging-up of
caving caving with outcropping or close to surface backfill. High stoping overlying rocks, displacement
frontal orebodies. Blind deposits with productivity. Mining in with and burst-like failure of large
(frontal-areal) unstable overlying rocks. Limited many stoping faces, extracted ore structural blocks. Forced
ore drawing application in gently dipping and quality control. Automation of ventilation of ore drawing-off
inclined deposits. Medium hard and stoping. Sufficient extraction workings (Figures 9a and
hard ores, non-liable to spontaneous index with frontal-areal ore 9b).
ignition and caking. Low and drawing. Ventilation due to mine
medium grade ores. Depth of depression.
application is to be determined.
TABLE 1 cont ...
Class Group Mining method Application conditions Advantages Disadvantages

3. Strike-line sublevel Thin and medium thick, Use of mobile machinery, High ore losses. High
stoping high dip orebodies. high productivity. dilution, inclusive of
Mining front divaricates. Automation. secondary dilution
Medium stable ores, (Figure 10).
stable host rocks. Low
and medium grade ores.
Depth of application is to
be determined.
4. Dip-line sublevel Thin and very thin, high Selective mining, without High labour intensity.
stoping dip orebodies, medium nonmetallic rocks, slot Great amount of
stable ores, stable host faces. Comparatively high preparatory driving. Use
rocks. Depth of ore extraction index. of portable equipment
application is to be (Figure 11).
determined.
III. Ground A. Mining with 1. Primary horizontal Thick and very thick, Overlying rock mass Mining in stages. Higher
control with alternating backfill and slicing with ascending gently dipping and stability. Lower costs as requirements to ore
backfill and caving (descending) mining and weakly inclined compared with Class I extraction standards.
caving paste backfill and orebodies, unstable and methods. Lower extraction index as
secondary block medium stable ores, compared with Class I
(sublevel) caving medium stable overlying methods (Figure 12).
rocks. Medium grade ore.
Depth of application is to
be determined.
2. Continuous room work Thick and very thick, Overlying rock mass Mining in stages. Higher
with primary mining with gently dipping and stability. Higher stoping requirements to ore
backfill and secondary weakly inclined productivity as compared extraction standards.
mining with block orebodies. Medium stable with the previous method. Lower extraction index as
(sublevel) caving. ores and host rocks. compared with Method 1
Medium grade ore. Depth in Class II.
of application is to be
determined.
6 7 Á-Á À Á À-À 2
3
7
1 2 5 8-12
4
À Á
FIG 1 - Class I, Group A, Type 1: 1 – inclined shaft; 2 – entry; 3 – levels; 4 – stoping; 5 – backfilling; 6 – level extraction;
7 – roof safety layer (please refer to Table 1).
8
Á-Á 7
1
2
À-À 4 3
À À
Á Á
5
1 3 6
FIG 2 - Class I, Group A, Type 2: 1 – inclined shaft; 2 – ore chute; 3 – cross adit; 4 – stopes; 5 – air raise; 6 – twin stope drifts;
7 – air-fill raise; 8 – filled levels (please refer to Table 1).
1 4
3
FIG 3 - Class I, Group A, Type 3: 1 – stope; 2 – soft pillars; 3 – backfill; 4 – orebody (please refer to Table 1).
2 4
3,5
8-12 8-12
5 3 1
FIG 4 - Class I, Group A, Type 4: 1 – entry; 2 – extraction levels in a chamber; 3 – temporary pillar (secondary chamber);
4 – orebody; 5 – filling mass (please refer to Table 1).
61
3 5
61
2
61
FIG 5 - Class I, Group A, Type 5: 1 – gallery; 2 – ore chute; 3 – langwall; 4 – scraper excavator; 5 – pack; 6 – raise (please refer to Table 1).
À Á-Á 3
À-À Á 3 7
5
3,5-4,5
8-10
1 2 4 8
8-12 1
À 2 Á
FIG 6 - Class I, Group B, Type 6: 1 – haulage heading; 2 – drilling and haulage roadway; 3 – air ort; 4 – ventilation connection;
5 and 6 – air drifts; 7 – orebody; 8 – filling mass (please refer to Table 1).
6 À Á-Á 4 5
Á À-À
2
2
1 3
9 À 10 1 7 Á
FIG 7 - Class I, Group B, Type 7: 1 – haulage heading; 2 – drilling roadway; 3 – entry; 4 – air and backfill cross-drifts; 5 – air and backfill ort;
6 – filling holes; 7 – raise; 8 – broken ore; 9 – filled primary chamber; 10 – secondary chamber (pillar) (please refer to Table 1).
À-À Á-Á
Á À
2
2
40
3
8
Á 50 À
FIG 8 - Class I, Group A, Type 1: 1 – fan-patterned holes intended to cave safety shield; 2 – main fan-patterned holes; 3 – drilling and
haulage level; 4 – ore chute (please refer to Table 1).
À-À Á- Á
5
13-14 ì
2 3 3
Á 15-20 ì Á
1
4
12-14 ì
FIG 9a - Class II, Group B, Type 2: sublevel caving with areal-frontal ore drawing: 1 – haulage roadway; 2 – drilling and haulage cross-drift;
3 – loading slopes; 4 – broken ore; 5 – air drift (please refer to Table 1).
Á-Á À-À
À Á
22-30ì
3
11-15ì
10-15
À Á
40-60
FIG 9b - Class II, Group B, Type 2: sublevel caving with frontal ore drawing: 1 – ore chute; 2 – incline shaft; 3 – drilling and haulage
cross-drift (please refer to Table 1).
A B
FIG 10 - Class II, Group B, Type 3: (A) mid-thick orebody; (B) thin orebody: 1 – haulage heading; 2 – subdrift; 3 – holes; 4 – under cantilever
area; 5 – crown pillar; 6 – caved rock; 7 – broken ore (please refer to Table 1).
REFERENCES
Agoshkov, M I, 1965. Calculating and Designing Ore Mining Methods
and Technologies (Nauka: Moscow).
Baikonurov, O A, 1969. Classification and Selection of the Underground
Mining Methods (Nauka: Alma-Ata).
Bronnikov, D M, Zamesov, N F and Bogdanov, G I, 1982. Deep Ore
Mining (Nedra: Moscow).
Freidin, A M, Neverov, S A and Neverov, A A, 2008. Mine stability with
application of sublevel caving schemes, Journal of Mining Science,
44:82-91.
Imenitov, V R, 1978. Underground Mining Methods for Ore Deposits
(Nedra: Moscow).
Khomyakov, V I, 1984. Foreign Experience of Mining with Backfill
(Nedra: Moscow).
Oparin, V N, Rusin, E P and Tapsiev, A P, 2007. World Experience
Gained in Underground Mining Automation (SO RAN: Novosibirsk).
Pilenkov, Yu Yu, Freidin, A M and Antipov, A V, 1990. Estimation of
natural impact of the origination of rockburst hazard at shallow
FIG 11 - Class II, Group B, Type 4: 1 – haulage heading; depths, in Rock Pressure and Underground Ore Mining Technology
2 – inclined shaft; 3 – ore chute; 4 – haulage cross-drift for Large Depths (IPKON AN SSSR: Moscow).
(please refer to Table 1). Trushkov, N I, 1947. Ore Mining, Volume II: Mining Methods
(Metallurgizdat: Moscow).
Yun, R B, Gerasimenko, V I and Malyshev, V N, 1997. Dynamic rock
Backfill
pressure manifestations at Zheskazgan deposit, Gorny Zhurnal, No 3.

Orebody Zamesov, N F, 1979. Influence of the ground control on the design of the
mining method for gentle ore formations, in Problems of
Underground Ore Mining at Large Depths (IPKON AN SSSR:
Moscow).

Caved

FIG 12 - Combined ground control; caving and backfill

(please refer to Table 1).
HOME
Resource Risk Quantification for a Deep Level Gold Mining

Operation
T Flitton1 and R Peattie2
ABSTRACT Moab Khotsong is wholly owned by AngloGold Ashanti and

has a published mineral resource of 13.79 million ounces of gold
South Africa’s Moab Khotsong Mine, one of the youngest mines to have
been brought into production in the historic Witwatersrand Basin, is and reserves of 6.97 million ounces. In addition to the gold
currently in its development phase, building up to full production around resource, 39 million pounds of uranium (17 723 contained
2010. It is located in a structurally complex and highly faulted piece of uranium tonnes) exist within the lease at a grade of 0.76 kg/t
ground, with the bulk of the known mineral resource lying between 2100 (AngloGold Ashanti Annual Report, 2007). Project Zaaiplaats 2
and 3700 m below surface. Moab Khotsong is wholly owned by (PZ2: the extension project) is aimed at optimally extracting the
AngloGold Ashanti Limited and has a published mineral resource of deeper portion (lower mine) of the Vaal Reef at Moab Khotsong,
13.79 million ounces of gold and reserves of 6.97 million ounces. In with the potential then existing at Moab Khotsong for the mine
addition to the gold resource, 39 million pounds of uranium exist within to operate until 2045. Investigations into extracting this block of
the lease at a grade of 0.76 kg/t. The recent extension of Moab ground have been underway for several years, but unfavourable
Khotsong’s mining life has received much attention, with the potential for
market conditions have delayed the initiation of mining in this
the mine to operate until 2045. This extension is acutely strategic for both
the mine and the company, since the life of the operation has increased deeper block of ground (up to 3500 m below surface).
significantly over the last few years as the exploration drilling continues Subsequent improvements in the world markets, not the least of
and the geological model evolves. A resource risk quantification exercise which has been the strengthening gold price, have lead
was undertaken in early 2008 to account for the sensitivity of Project AngloGold Ashanti to revisit the deepening of Moab Khotsong’s
Zaaiplaats 2 to the risks identified in an independent technical review. operations. A prefeasibility study on the viability of the project
This resource risk quantification exercise involved several scenarios was recently completed and Project Zaaiplaats 2 is currently in
which tested the sensitivity of gold values – and thus gold content – to the feasibility phase. Although the prefeasibility study shows a
changes in the geological facies and structural modelling. This paper will relatively low return on investment, Project Zaaiplaats 2 is being
briefly introduce the process followed for quantifying uncertainty in the viewed in terms of the potential this project has of increasing the
geological facies and the structural modelling, then present the results of
life of the entire Vaal River operations (AngloGold Ashanti
this exercise. The paper will highlight the potential financial impact that
this quantified change could have on the operation and demonstrate how Annual Report, 2007; de Beer et al, 2007). As a result, the
the geological risk has been incorporated into the financial modelling. A AngloAmerican Group’s Anglo Technical Division was
summary and comparison of the results for each scenario will be requested by AngloGold Ashanti to conduct a review of the
provided, as well as suggestions for further understanding the resource Project Zaaiplaats 2 prefeasibility study (Faul et al, 2007). A
uncertainty and quantifying its potential impact on the project. number of recommendations, which were to be implemented prior
to feasibility, were made by Anglo Technical Division to address
INTRODUCTION some elements of the identified high-risks regarding geology and
mineral resources. This paper summarises the sensitivity analysis
The AngloGold Ashanti South African operations comprise that was undertaken in early 2008 to account for the uncertainty
seven underground mines which are located in two geographical regarding geology and mineral resources identified in the
regions in the Witwatersrand Supergroup (deposited in the area independent technical review – it ameliorates the risk involved
often described as the Witwatersrand Basin) – known as the with palinspastic reconstruction and interpretation of the
Carletonville (West Wits operations) and Klerksdorp (Vaal River geological facies.
operations) gold fields (Figure 1 – left) (Anglogold Ashanti
Annual Report, 2007). Moab Khotsong Mine, the youngest of
AngloGold Ashanti’s South African operations (Figure 1), began GEOLOGICAL SETTING
commercial production in January 2006 and is planned to be at AngloGold Ashanti’s South African operations are all located in
full production around 2010. The mineral resource at Moab
the Witwatersrand Basin, regarded as one of the greatest
Khotsong is structurally complex and highly faulted, with large
metallogenic provinces of the world (Robb and Robb, 1998). The
fault-loss areas and the bulk of the known mineral resource lying
between 2100 and 3700 m below surface (AngloGold Ashanti Witwatersrand Supergroup (2.7 Ga to 2.9 Ga) comprises a 6 km
Annual Report, 2007). It is situated south and south-east of Great thick sequence of predominantly argillaceous and arenaceous
Noligwa and Kopanang in the Klerksdorp (Vaal River) gold field, sediments that extend laterally for some 300 km north-east/south
in the Free State Province (AngloGold Ashanti Annual Report, -west and 100 km north-west/south-east on the Kaapvaal Craton.
2007). The principal reef is the Vaal Reef, with mining grades of The upper portion of the sequence (the Central Rand Group)
between 10 g/t and 20 g/t gold. The Vaal Reef is commonly contains the laterally extensive, economically mineralised, gold-
represented by a series of oligomictic conglomerates and bearing quartz pebble conglomerate horizons or reefs (Robb and
quartzite packages developed on successive non-conformities. Robb, 1998). Only a small portion of the Witwatersrand Super-
Several distinct facies have been identified (Figure 2), each with group is exposed in outcrop, most being overlain, unconformably,
unique geological characteristics, gold distributions and grade by rocks of the Venterdorp (2.7 Ga), Transvaal (2.6 Ga) and Karoo
characteristics. (2.80 Ma) Supergroups (Robb and Robb, 1998; McCarthy, 2006).
The reefs, generally less than two metres thick, are widely
considered to represent laterally extensive braid-plain and lesser
1. Geoscience Manager, AngloGold Ashanti Limited, Private Bag alluvial fan environments (Robb and Robb, 1998). The highly
X2051, Carletonville 2500, South Africa. faulted Klerksdorp gold field (Figure 1) lies to the southwest of
Email: [email protected] the Caletonville gold field and is separated by the Potchefstroom
2. MAusIMM, Vice President, Mineral Resources Management, Gap. The geology of the area is complicated by erosional
AngloGold Ashanti Limited, PO Box 62117, Mashaltown 2107, truncation of lower formations and overstepping by younger
South Africa. Email: [email protected] formations (Robb and Robb, 1998). There is still much debate
T FLITTON and R PEATTIE
FIG 1 - Location map of the Witwatersrand Basin in South Africa, with younger cover sequences removed (left), showing the Vaal River
operations (right).
Moab Khotsong Mine

Great Noligwa Mine
Kopanang Mine
Facies interpretation
at pre-feasibility
Great Noligwa Mine Great Noligwa Mine

Kopanang Mine Kopanang Mine
Scenario D Scenario E
optimistic 500 pessimistic 500
Facies Facies
FIG 2 - Varying facies interpretations – facies interpretation at prefeasibility (top) versus scenario D (bottom left) and scenario E
(bottom right).
RESOURCE RISK QUANTIFICATION FOR A DEEP LEVEL GOLD MINING OPERATION
about the origin of the gold mineralisation in the Witwatersrand poorly-explored areas (de Beer et al, 2007). Using the mapped
Basin. Gold was generally considered to have been deposited geological characteristics of the conglomerates of the A and C
syngenetically with the conglomerates, but increasingly an horizons, the Kopanang-Great Noligwa-Moab Khotsong mining
epigenetic theory of origin is being supported (Robb and Robb, area is currently subdivided into nine geozones, which are
1998). Nonetheless, the mineralisation is invariably associated viewed as geologically homogenous regions suitable for
with erosional disconformities within the sequence in various estimation (de Beer et al, 2007). Only four of these geozones
host lithologies. Gold generally occurs with pyrite, uraninite and (500/370/440/460) are present within the Moab Khotsong Lower
carbon, with quartz being the main gangue mineral (Robb and Mine (Figure 2). The 470 geozone currently present in Great
Robb, 1998). Noligwa Mine and Kopanang Mine between the 500 and 440
geozones is not present in the Moab Khotsong Lower Mine.
PROJECT ZAAIPLAATS 2 MINERAL RESOURCE
Structural model
Facies model
The structural component of the geological model was
The facies model (Figure 2) within Project Zaaiplaats 2 area is synthesised from the analysis of vibroseis survey data, along
based largely on 15 surface boreholes (multiple reef intersections with the structural information projected from contiguous mining
per borehole) and prior knowledge from the Great Noligwa and areas. Important stratigraphic and structural information was then
Kopanang Mines to the north (Figure 3). Experience has shown derived from the detailed analysis of the 15 surface boreholes and
that an understanding of the processes that led to the formation their deflections that intersected the area, and used together with
of the reef and the distribution of gold within that orebody is of vibroseis interpretation, to identify the major faults in the area.
crucial importance for estimation to be as accurate as possible These major faults were then used to define the margins of the
(de Beer et al, 2007). The Vaal Reef package on Great Noligwa, Vaal Reef blocks and the expected Vaal Reef horizon was then
Kopanang and Moab Khotsong Mines has been shown to consist constructed by projection parallel to the structural dip as far as
of three correlatable conglomerate units – the ‘A’, ‘B’ and ‘C’ the major faults (de Beer et al, 2007). The resulting structural
members. Changes in the nature and inter-relationships between architecture was tested by reconstruction. In this process, all reef
these units are reflected in the broad-scale distribution of gold on blocks are palinspastically restored to their original unfaulted
Great Noligwa and Kopanang. Such changes are often subtle and locations prior to movement on the Jersey Fault (movement
require detailed logging and correlation work. The geological modelled to be 1500 m to the south-east), adjacent to the Great
characteristics that define the reef in each of the areal Noligwa and Kopanang mines (Figure 3). By restoring the reef
subdivisions (geozones) are then compared to counterparts in blocks to their interpreted unfaulted locations, it becomes
borehole core, and the ‘Facies Model’ is extended into possible to interpret how the different units of the Vaal Reef,
Kopanang Mine
Moab Khotsong Mine
Great Noligwa Mine
±1
50
0m
Reconstruction
at pre-feasibility
Great Noligwa Mine

Kopanang Mine Kopanang Mine
Scenario B Scenario C
rec onstruction reconstruction
±430m further west ±640m further east
FIG 3 - Varying palinspastic reconstruction – reconstruction at prefeasibility (top) versus scenario B (bottom left) and scenario C
(bottom right).
intersected in the surface drilling from the Moab Khotsong and relevance of the geological model. This is particularly the case
blocks, correspond with the projected facies plans from the in a Bayesian geostatistical approach, and even more so when
adjacent Great Noligwa and Kopanang properties (de Beer et al, reconstruction of data has taken place as in project Zaaiplaats.
2007). In addition to the structural blocks thus identified,
extensive geological target areas are known to exist, but at a RESOURCE RISK QUANTIFICATION
poorly-defined elevation. Exploration drilling in these blocks is
likely to realise the upside potential of the project (shown in The structural modelling (palinspastic reconstruction) and the
Figure 3 as the hatched reef blocks). Geological target areas do geological facies were identified as high-risks in the Independent
not form part of the published mineral resource (de Beer et al, Technical Review Report (Faul et al, 2007). They suggested the
2007). following course of action for risk mitigation:
The reconstruction of the blocks to a pre-Jersey
Estimation methodology Fault position needs to be tested as it is indicated
The sampling data for Project Zaaiplaats is comprised of that there is an alternative to the current model
underground chip samples, underground boreholes and surface with a smaller sinistral component.
boreholes, with the surface boreholes being of a different support The facies boundaries used for the current
to the underground boreholes and chip sampling. All sample reconstruction of a pre-Jersey fault situation are
locations are reported as a composite over a mineralised width, soft boundaries, transitions rather than hard
resulting in a single channel width (cm) and metal accumulation boundaries. Therefore the interpretation of the
(cm g/t) value. facies in the main eastern block of the project
AngloGold Ashanti makes use of a Bayesian geostatistical needs to be tested with borehole intersections as
approach for resource estimation at project Zaaiplaats 2, where the the knowledge of the possible presence or absence
absence of dense sampling data gold estimation is based on a of the barren 470 facies is of economic
combination of the observed data and external knowledge relating importance. In the current model the 470 facies
to the data. A Bayesian geostatistical approach asserts that the area does not occur in the project area, however, it
to be evaluated forms part of a larger continuous entity, to which may still occur in the place of the current
the observed data belongs. The assumption behind this approach is 440 facies.
that the geological, statistical and spatial characteristics observed
in the known areas hold for the ‘unknown’ uninformed areas. If The resource risk quantification exercise was undertaken to
the assumption is valid, then the orebody can be evaluated over ameliorate the risk involved with palinspastic reconstruction and
large uninformed areas, where limited information exists. Geozone interpretation of the geological facies. The resource risk
constraints are critically important within this process, because the quantification exercise involved the following scenarios, testing
geozones subdivide the data into distinct populations and the the sensitivity of values, ie gold, to changes in the geological
parameters of these populations play a critical role in the facies and reconstruction:
development of the estimates (Chamberlain, 1997).
Mixed support cokriging was used in the estimation of the A. Moab Khotsong Lower and Upper Mine data in
mineral resource for the project. It is a technique that enables the unconstructed space, rekriged and re-estimated.
use of data of mixed support, allowing both borehole and B. Moab Khotsong Lower Mine reconstruction further west
underground sampling data to be used together (Chamberlain, compared to reconstruction used in prefeasibility, rekriged
1997). Project Zaaiplaats is currently estimated as part of a larger and re-estimated.
entity, recognising that the Vaal Reef is present over a number of
mines. The models that are developed are therefore on a regional C. Moab Khotsong Lower Mine reconstruction further east
scale, rather than a mine scale. Estimation is performed into 300 m compared to reconstruction used in prefeasibility, rekriged
× 300 m sized blocks, which fully capture the within block and re-estimated PZ2 with more optimistic view on 500
variance, allowing the cokriging of data of different support sizes facies.
over long ranges. The variography is based on the combination of D. Moab Khotsong Lower Mine geological facies having an
the block variograms and borehole variograms (Chamberlain, optimistic outlook on 500 facies.
1997). Estimation is done in log space because of the highly
skewed gold distribution found within the Vaal Reef. The final E. Moab Khotsong Lower Mine geological facies having a
gold estimates are then calculated by back transforming the kriged pessimistic outlook on 500 facies.
estimates, using the compound log normal (CLN) four parameter F. Moab Khotsong Lower and Upper, Great Noligwa Mine
distribution models (Chamberlain, 1997). The current method has and Kopanang Mines having no facies boundaries (data
proven reliable (resource estimates derived from mixed support
combined), rekriged and re-estimated.
cokriging have been compared against production results on well
established mines in the same area) and has been in use on the G. Different global estimation methods.
Witwatersrand reefs, particularly the Vaal Reef, for a number of
years. The mixed support cokriging global estimation method has Scenario A
undergone intense scrutiny by a number of external auditors,
including AngloAmerican Group’s Anglo Technical Division, Scenario A is intended to quantify the sensitivity on Project
Snowden and SRK. Zaaiplaats 2 if the faulted blocks were not reconstructed at all.
Due to the limited amount of borehole data within the project The data (samples from boreholes and/or chip samples) as well
area (15 mother surface boreholes) and the distance of the project as the structure would remain in true space for Moab Khotsong
to surrounding mines (less than one kilometre), the Moab Upper and Lower Mines (Figure 3).
Khotsong Lower Mine reef blocks have been reconstructed back
to their original unfaulted positions, prior to movement on the Scenario B
Jersey Fault, adjacent to Great Noligwa and Kopanang mines
(Figure 3). There is an implicit risk in this, in that the ‘true’ Scenario B is intended to quantify the sensitivity on Project
movement vectors could be different to what is currently used, but Zaaiplaats 2 if the faulted blocks were reconstructed a further
the benefit in terms of additional sampling information being 430 m west from the reconstruction used in the prefeasibility
available potentially outweighs the risk. Any estimation process is (movement modelled to be 1500 m to the south-east). The data
largely reliant on the validity of the data as well as the accuracy (samples from boreholes and/or chip samples), as well as the
RESOURCE RISK QUANTIFICATION FOR A DEEP LEVEL GOLD MINING OPERATION
TABLE 1
Comparison of results from the resource risk quantification exercise.
Project Zaaiplaats 2 Au (Moz) Diff in Au (Moz) % diff in Au (Mz)

Feasibility 7.683
Scen A – No palinspastic reconstruction 6.932 -0.751 -10.84%
Scen B – Palinspastic reconstruction further west 7.090 -0.593 -8.37%
Scen C – Palinspastic reconstruction further east 7.566 -0.117 -1.54%
Scen D – Optimistic 500 facies 7.796 0.113 1.44%
Scen E – Pessinistic 500 facies 7.525 -0.158 -2.10%
Scen F – No facies boundary (max samples 30) 8.033 0.350 4.35%
Scen G – Average value 6.187 -1.496 -24.18%
Scen G – CLN estimate using chip samples variances 7.694 0.011 0.14%
Scen G – CLN estimate using surface borehole variances 6.390 -1.293 -20.23%
Scen G – Sichel – T estimate using chip samples variances 8.139 0.456 5.60%
Scen G – Sichel – T estimate using surface borehole variances 6.378 -1.305 -20.45%
Scen G – Two parameter lognormal estimate 6.749 -0.934 -13.84%
structure of the Moab Khotsong Lower Mine faulted blocks, doing the estimation. This scenario will address the absence of
would be reconstructed a further 430 m west of the re- the 470 reef facies at Moab Khotsong Lower Mine. All data is
construction used at prefeasibility. Moab Khotsong Upper Mine merged (including the 470 facies) into one single dataset and
would maintain its current interpretation of reconstruction as estimated accordingly. All the data (samples from boreholes and/
used in the prefeasibility, as it is a different faulting system or chip samples), as well as the structure of the Moab Khotsong
compared to the Jersey Fault (Figure 3). Upper and Lower Mine faulted blocks, would be reconstructed
using the reconstruction vectors at prefeasibility.
Scenario C
Scenario C, like scenario B, is intended to quantify the sensitivity Scenario G
on Project Zaaiplaats 2 if the faulted blocks were reconstructed a Various non-spatial global estimation methods were used to
further 640 m east of the reconstruction used during prefeasibility. calculate global estimates for various portions of the Moab
The data (samples from boreholes and/or chip samples), as well as Khotsong Lower Mine, thus obtaining a range of values for the
the structure of the Moab Khotsong Lower Mine faulted blocks, Project Zaaiplaats 2 area. The methods included arithmetic
would be reconstructed a further 640 m east of the reconstruction average of the surface boreholes, a two parameter lognormal
used at prefeasibility. As with scenario B, all the data as well as
estimate using value of surface boreholes, Sichel-T estimates
the structure of the Moab Khotsong Upper Mine faulted blocks,
would only be reconstructed using the reconstruction vectors at using surface boreholes, CLN estimates using surface boreholes.
prefeasibility (Figure 3). All of these estimates are further compared to the current
prefeasibility estimates for Project Zaaiplaats 2.
Scenario D
SUMMARY AND COMPARISON OF RESULTS
Scenario D is intended to quantify the sensitivity on Project
Zaaiplaats 2 if the facies lines were adjusted, whilst still keeping It was noted that for scenarios A, B and C, where reconstruction
the true facies interpretation of the surface boreholes. A more movements were varied (Figure 3), the new mineral resource
optimistic interpretation of the 500 reef facies is used in place of estimates all decreased (Table 1). In the scenario where no re-
the facies interpretation at prefeasibility (Figure 2). All the data construction was used (A) the highest difference in value
(samples from boreholes and/or chip samples), as well as the compared to the prefeasibility mineral resource was calculated.
structure of the Moab Khotsong Upper and Lower Mine faulted Scenario A is not considered as a realistic alternative based on the
blocks, would be reconstructed using the reconstruction vectors at current understanding of the geology, however, it does give a
prefeasibility.
worst-case scenario if either the reconstruction vectors utilised are
inaccurate, or if the mineralisation was post the final fault
Scenario E movement. Scenario B is therefore considered as the most likely
Scenario E, like scenario D, is intended to quantify the sensitivity worst-case scenario (highest change in value from prefeasibility) –
on Project Zaaiplaats 2 if the facies lines were adjusted, whilst still the potential financial impact this quantified change could have on
keeping the true facies interpretation of the surface boreholes. A the operation is shown in Table 2.
more pessimistic interpretation of the 500 reef facies is used in Scenarios D and E, where facies interpretation was varied, while
place of the facies interpretation at prefeasibility (Figure 2). All honouring the surface boreholes logged facies interpretation as
the data (samples from boreholes and/or chip samples), as well as used in the prefeasibility, showed negligible difference compared
the structure of the Moab Khotsong Upper and Lower Mine
to the prefeasibility mineral resource estimate (Table 1). While for
faulted blocks, would be reconstructed using the reconstruction
vectors at prefeasibility. scenario F, where no geological modelling was used in the
estimation, the estimates were higher than those calculated for the
prefeasibility. The relationship of variance for the combined data,
Scenario F compared to each facies individually, resulted in more weighting
Scenario F is intended to quantify the sensitivity on Project given to closer (higher value) regularised blocks in this scenario.
Zaaiplaats 2 if no facies lines (ie hard boundaries and similar Scenario G involved the calculation using various global
geological facies characteristics are ignored) are used when estimation methods, resulting in a range of values for various
TABLE 2 values to the south (together with the reconstruction vectors at pre-
feasibility of 1500 m north-west from faulted true positions),
Effect on IRR (standardised) of PZ2 if scenario B (worst-case
scenario – further west palinspastic reconstruction) is used as the
should be completed to potentially quantify the lack of presence of
likely reconstruction. this facies. Both scenarios are to be completed in the near future,
thus fully quantifying the sensitivity on Project Zaaiplaats 2.
Prefeasibility Scenario B Percentage
difference
REFERENCES
On reef value (cm g/t) 2597 2401 -8%
AngloGold Ashanti Annual Report, 2007. Supplementary information:
Standardised IRR (%) 1 0.75 -25% Mineral resources and ore reserves, p 124.
Chamberlain, V A, 1997. The application of macro co-kriging and
compound lognormal theory to long range grade forecasts for the
portions of the Moab Khotsong Lower Mine. The majority of the Carbon Leader Reef, project report submitted in partial fulfillment of
global estimates calculated in scenario G are below the the requirements of Master of Science in Engineering (unpublished),
prefeasibility Project Zaaiplaats 2 published value estimate. This University of the Witwatersrand.
scenario highlights that Project Zaaiplaats 2 is very sensitive to the De Beer, R, McIvor, J, Broodryk, A, Kotze, B, Uys, J, Engelbrecht, R,
Flitton, T M, Frith, I, Botha, L and Charalambous, V, 2007. Africa
amount of the project area that is covered by the 370 facies. underground region, Moab Khotsong Mine, business plan 2008: R3,
Project Zaaiplaats 2 has a high published resource value estimate, Bp 15 Book, Project Zaaiplaats 2, study (unpublished).
mainly because the bulk of the ground consists of the ‘highest Faul, H, Rompel, A K K, Rice, P, Carvil, P, Roman, N, Meyer, G,
value’ 370 reef facies. Thus a very real risk to the extension Stephenson, J D, van der Linde, A, Hacking, T, Hutton, C, Johnson,
project exists if there is not as much 370 reef facies as is currently A, Magqabi, L, Oldfield, R and Cantrell, K, 2007. Independent
modelled, in which case the value can only decrease. technical review report on behalf of the AAplc group technical
director of the AngloGold Ashanti Ltd Moab Khotsong Mine PZ2
Zaaiplaats prefeasibility study (unpublished).
RECOMMENDATIONS FOR FURTHER STUDY McCarthy, T S, 2006. The Witwatersrand Supergroup, in The Geology of
Two additional scenarios are to be completed in the resource risk South Africa (eds: M R Johnson, C R Anhouser and R J Thomas),
quantification exercise. A multi-gaussian conditional sequential pp 155-186 (Geological Society of South Africa, Council for
Geoscience: Pretoria).
simulation is suggested as a means to get a range of estimates
Robb, L J and Robb, V M, 1998. Gold in the Witwatersrand Basin, in The
located in space for the project area. A final scenario where the
Mineral Resources of South Africa Handbook (eds: M G C Wilson
470 facies is shown to be present in the Moab Khotsong Lower and C R Anhouser), 16:294-349 (Council for Geoscience).
Mine, using the data from the 470 facies in the north to predict the
HOME
An Application of Genetic Algorithms to the Main Ventilation

System Optimisation for Multi-Level Mine Operations –
The Impact of the Fitness Function
E Acuña1, S Hardcastle2,3, L Fava4,5, P Dunn6 and S Hall7
ABSTRACT airflow from the surface to the vicinity of mining areas, then
back to the surface. The secondary (or auxiliary) ventilation
An operational ventilation network has to deliver the airflow required for
the planned activities in an underground environment. The objective in
system is responsible for creating an airflow from the primary
designing a ventilation network is to select the fan locations and system to the dead-ends where mining development and
operation points that will provide the required pressures and airflows production generally occur. The main purposes of ventilation are
while achieving minimum or near-minimum cost. Historically, the to bring fresh air (oxygen) to the different areas in the
methodology for ventilation optimisation has been based on case studies underground environment for people and machines, then to dilute
and expert knowledge. Additionally, some more rigorous mathematical and remove contaminants generated by the mining process.
approaches have been developed, and applied to different networks. The Adequate ventilation is one of the key conditions enabling the
optimisation of primary mine ventilation systems using genetic mine to safely perform the activities scheduled by mine planners.
algorithms has been demonstrated under two approaches: with a fitness Considering the operation of fans, along with heating and
function based on airpower consumption, and with a fitness function cooling systems to condition the air, ventilation can account for 15
based on the deviation of airflow from the requirements. Unfortunately, to 22 per cent of a mine’s operating cost, as stated by Petrov and
the different approaches were applied to different networks, making it
Popov (2004). However, from an energy usage perspective,
difficult to compare the benefits and tradeoffs of the fitness functions.
The main objective of the proposed approaches is to improve the value of
ventilation costs in mechanised mines can account for around
mining plans by accounting for ventilation requirements, where the 40 per cent of the consumption, according to Hardcastle and
improvement can be realised by two means: first, by allowing the Kocsis (2007); the proportion of energy usage by primary,
schedules to be executed as planned, reducing (or avoiding) development secondary, heating and cooling can vary considerably as stated by
and production losses, and second, by reducing ventilation costs. This Hardcastle et al (2007). Consequently, the energy consumption of
paper presents a comparison between the airpower and the airflow ventilation systems is becoming one of the most important cost
deviation fitness functions to improve main mine ventilation systems. An control areas within underground mines. With increasing pressure
application to a mine ventilation network is presented and solved with on mines to reduce energy consumption and greenhouse gas
both fitness functions, and then compared to determine the advantages (GHG) emissions, mining operations are being forced to develop
and disadvantages taking into account the airpower and airflow deviation new approaches and design tools, and also to incorporate the latest
values obtained for the network and the solving time. mining technologies in order to lower production costs while
meeting production and environmental requirements, according to
INTRODUCTION Kocsis, Hall and Hardcastle (2003).
Currently, the optimisation of mine ventilation systems is
An adequate mine ventilation system is one of the key typically a manual decision process. It is assisted by computers,
components to allow development and production in using commercial ventilation distribution solvers (referred to as
underground environments. Generally, the mine ventilation network simulators within the mining industry) to test alternatives
system can be divided into two levels or functions: primary and proposed by the ventilation expert. From those alternatives, the
secondary. The primary (or main) ventilation system, containing one that reports the lowest energy consumption value while
the main and booster fans, is solely responsible for generating an satisfying the peak airflow requirements is typically selected and
called the optimal solution. Mine ventilation optimisation based
on ventilation solvers is accepted by the mining industry as a
1. PhD Candidate, Natural Resources Engineering, MIRARCO, reliable methodology to decide good feasible alternatives to
Laurentian University, 935 Ramsey Lake Road, Sudbury Ontario implement in mine ventilation networks. The decision as to
P3E2C6, Canada. Email: [email protected] whether or not such an optimised solution is indeed the optimal
2. Adjunct Professor, School of Engineering, Laurentian University, solution is dependent on the input criteria, and how well it
935 Ramsey Lake Road, Sudbury Ontario P3E2C6, Canada. represents the actual conditions within the mine over time.
Email: [email protected] Because ventilation requirements can fluctuate significantly in the
short term, while overall demand changes gradually over the long
3. Senior Research Scientist – Ventilation Specialist, Mine Air Quality
and Ventilation, CANMET, Natural Resources Canada, 1079 Kelly term, these manually selected solutions could be suboptimal. To
Lake Road, Sudbury Ontario P3E5P5, Canada. date, various attempts have been made (Lowndes, Fogarty and
Email: [email protected] Yang, 2005; Lowndes and Yang, 2004; Calizaya, McPherson
and Mousset-Jones, 1987; Wu and Topuz, 1998) to generate a
4. Adjunct Professor, Department of Mathematics and Computer mathematical mine ventilation optimisation technique or process
Science, Laurentian University, 935 Ramsey Lake Road, Sudbury and have concentrated on a single time period and its optimal fan
Ontario P3E2C6, Canada. Email: [email protected]
selection. Aside from Lowndes and Yang (2004) and Acuña et al
5. Research and Development Manager, Symboticware Incorporated, (2009), mine ventilation optimisation techniques have not been
83 Durham Street, Sudbury Ontario P3E3M5, Canada. implemented across multiple periods.
6. Director – Western Australian School of Mines, Faculty of Science BACKGROUND
and Engineering, Curtin University of Technology, GPO Box U1987,
Perth WA 6000. Email: [email protected]
Mine ventilation network optimisation
7. President and CEO/Professor, School of Engineering, MIRARCO,
Laurentian University, 935 Ramsey Lake Road, Sudbury Ontario Mine ventilation networks are modelled as sets of connected
P3E2C6, Canada. Email: [email protected] nodes and edges – these are respectively referred to as junctions
E ACUÑA et al
and branches by mine ventilation design practitioners. The edges Rj is the resistance factor for branch j
represent the different paths that the airflow can follow; the
nodes are the connections between those paths. Each edge has an HRj is the pressure drop of the regulator in branch j
associated resistance value that represents the friction that has to HFj is the fan pressure in branch j
be overcome by the airflow in order to traverse the edge. Because
of this friction, a pressure drop is generated in every branch HNj is the natural pressure in branch j
traversed by airflow. Fans are used to supply the pressure in m is the number of chords (meshes) in the network (m = b – n
order for the airflow to overcome the pressure drop of the + 1), bij = 1 if branch j is contained in mesh i and has the
different branches, and to deliver the right amount of airflow same direction, bij = -1 if branch j is contained in mesh i
through the ventilation network. Mine ventilation network and has the opposite direction, and bij = 0 if branch j is not
calculations are based on the network designs and Kirchhoff’s contained in mesh i
first and second laws, as stated by Lowndes, Fogarty and
Yang (2005) and McPherson (1993). Mine ventilation network L j ≤ Qj ≤ U j ∀j ∈ {1,..., b} (6)
optimisation problems are nonlinear and nonconvex according to
Xing and Topuz (1998). This means that no known classical
HR j ≥ 0 ∀j ∈ {1,..., b} (7)
optimisation method can solve this problem and guarantee an
optimum. Moreover, there is no known function to transform the
pressure distribution (vector P) into airflows in the mine where:
ventilation network. Iterative or optimisation methods have to be Lj and Uj are the lower and upper bound for the airflow in each
used to generate the airflow distribution. In addition, depending branch of the network and HRj is the additional resistance that
on the problem, the solution space for single and multiple period can be installed in each branch.
network problems can be large enough to make it impossible to The objective function and the pressure balance constraints
find the optimum solution in a reasonable time. (Kirchhoff’s second law) are nonlinear. The airflow balance
Mine ventilation network optimisation in terms of energy is constraints (Kirchhoff’s first law) are linear constraints. Airflow
written as a nonlinear optimisation problem of the following requirements as well as lower and upper bounds for the airflow
form: in the different branches can be added. A vector P that satisfies
all the constraints is a feasible solution of the problem. The
b
Pj ⋅ Qj
Min f ( x ) = ∑ 24 ⋅ Days ⋅ ⋅ EC feasible region F is the set of all feasible points. The objective of
j=1 nj ⋅1000 (1) the optimisation process is to find the best feasible P* ∈ F such
that f(P*) = f(P) for all P ∈ F. Vector P* is called the optimum
P = ( P1 ,..., Pb )t ∈ F ⊆ S ⊆ R b solution of the problem. In practical terms, the vector P* is the
set of pressure values for the fans that will deliver at least the
where:
required airflow for the mine, and at minimum cost.
b is the number of branches in the network In this study, two different fitness functions are implemented
Days is the number of days that the fan is going to be working, and compared: one based on the airpower and the other one on
usually on a yearly basis, ie, Days = 365; 24 is used to the square deviations from the airflow requirements. The idea is
transform the days into hours; EC is the energy cost to study the capability of these fitness functions to explore the
($/kWh); and 1000 transforms W into kW because of the solution space and find good feasible solutions or near optimal
EC units. P = (P1, … , Pb) is the vector of pressures, Q = solutions. The principal assumption is that a solution with
(Q1, … , Qb) is the vector of airflow resulting from the reduced airflow volumes will be less expensive in terms of
pressures, ç = (ç 1, … , ç b) is the vector of efficiencies for a airpower or energy cost, so it is expected that both fitness
combination of pressure and airflow if a fan is located in functions should converge to the same optimal or near optimal
the branch j; F is the feasible region, and S is the whole values.
search space In order to solve these equations, nonlinear optimisation or
iterative methods have been successfully implemented (Wu and
subject to:
Topuz, 1998; Calizaya, McPherson and Mousset-Jones, 1987;
b Lowndes and Yang, 2004). The most popular technique used to
∑a ij ⋅Qj = 0 ∀i = 2,..., n (2) solve mine ventilation networks, ie, find the airflows of the
network based on the pressures of the different fans, is the Hardy
j=1
Cross method. This method is based on an initial solution for the
where: ventilation network, and then converges to the real values of the
pressure drop and airflow on each of the branches of the network
Qj is the airflow quantity through branch j, n is the number of through an iterative process. The method is popular because for
nodes in the network, aij = 1 if branch j is connected to node i and most of the ventilation networks, it converges in a small number
the airflow goes away from node i, aij = -1 if branch j is of iterations, as stated by McPherson (1993). It uses a
connected to node i and the airflow goes into node i, aij = 0 if simplification of Kirchhoff’s first law; the method assumes the
branch j is not connected to node i; densities to be the same in all junctions. This is called the
incompressible approach, and the mass flow conservation is
b
transformed into airflow conservation. The incompressible
∑b ij ⋅ H j = 0 ∀i = 2,..., m (3)
approach is only valid for shallow mines to a depth of 500 m. For
j=1
deeper mines, the compressible approach is recommended.
H j = HL j + HR j − HFj − HN j ∀j ∈ {1,... b} (4)
Practical mine ventilation network optimisation
HLj = Rj ⋅ Q 2
j ∀j ∈ {1,..., b} (5) Several mine ventilation solvers are available on the market as
software tools. They are called mine ventilation simulation tools,
where: although they don’t perform any simulation process in the sense
of random variable distributions and discrete or continuous event
HLj is the friction pressure drop for branch j simulation. These tools solve the network and calculate the
AN APPLICATION OF GENETIC ALGORITHMS TO THE MAIN VENTILATION SYSTEM OPTIMISATION
values of the airflow in the ventilation networks. This is why Generate n offspring V’ by crossover of selected pairs and
they are known as ventilation network simulators, because they mutation
are able to ‘simulate’ (calculate the value of) the airflows. These
software tools have a network solver connected to a user (Crossover rate 0.6 and Mutation rate 0.1)
interface to display the network and the analyses available. Evaluate individuals of V’ with 3D-CANVENT ventilation
In order to solve mine ventilation networks, the Hardy Cross solver
technique is used to calculate the airflows in the network, based Calculate feasibility value, penalty and fitness of V‘
on the locations of the fans in the branches, resistances of the Let V = n best individuals from V and V’
branches, and the duties of those fans (fixed pressure or
pressure-airflow curves). The ventilation engineer can manually Return best solution (individual) found, and feasibility value
interact with the user interface to change the network parameters, End
or the fan locations and duties, and generate a new result for each
scenario, but the solver will only generate the solution for that CASE STUDY
particular scenario. Ventilation solvers have no capacity to
propose fan locations and duties in order to have a feasible The case study is based on the mine ventilation network of the
solution, or the best fan locations and duties to meet the airflow former Gedling Colliery coal mine in the United Kingdom. The
requirements and generate a feasible, economical solution. Based mine has one fresh air intake shaft and one exhaust shaft, with a
on the trial of several different combinations of the fan locations main surface exhaust ventilation fan system (MF), as stated by
and duties, using mine ventilation solvers (with the Hardy Cross Hardcastle (1983). A simplified schematic of the ventilation
technique), the best solution is identified and taken as optimal. network is displayed in Figure 1. There are five work faces or
production areas, labelled WF1 through WF5 on the diagram;
Constrained genetic algorithms WF1 and WF5 are simplifications of more extensive parts of the
ventilation system with multiple work faces. The airflow
Constrained genetic algorithms have been successfully used to
solve one period, large mine ventilation network problems, as requirements are specified for each area, and they depend on the
stated by Lowndes, Fogarty and Yang (2005) and Acuña et al type of activities taking place. In addition to the main surface fan,
(2009), using a penalty function with a large preset value to there are two booster fan systems (F1 and F2) required to assist
preserve feasibility. A constrained genetic algorithm doesn’t the surface fan duties in order to provide the airflow requirements
explore the entire solution space (S) with optimality as its prime through the mine. An additional booster fan is located in the
target, then project to the feasible region; instead, it explores the network to simulate the natural ventilation pressure.
space of feasible solutions (F) to find the best feasible solution There are three main objectives assigned to the GA decision:
possible. For this reason, the use of penalty functions can the number of fans to install in the mine, the fixed pressure of the
increase the likelihood that the final solution of the genetic fans (that will determine the airflows generated by those fans,
algorithm will be just a local optimum. Also, depending on the usually duties) and the location of the fans given the available
solution space and the operators of the genetic algorithm, the places within the mine ventilation network. The final decision
final solution can be infeasible, although this outcome may be has then three parts, ie (1) number (2) duties and (3) locations,
unlikely. Considering the financial relationship between for the network. The decision is the value of the fixed pressure to
production, development and ventilation, the foremost goal will be applied in a given location in the network. The number of
always be to have a feasible solution for ventilation to support locations with a fixed non-zero pressure value is the number of
the planned schedule, though the additional benefit of an fans that will be needed in the mine (decision 1). The duties are
improved solution will certainly be well received. the combination of the fixed pressures and the airflow resulting
from the fixed pressure combination (decision 2). Finally, the
GENETIC ALGORITHM IMPLEMENTATION locations will be those with non-zero pressure values (decision 3).
This study does not claim that the GA will find the optimal
Genetic algorithms (GAs) are search algorithms inspired by solution to the mine ventilation network optimisation; in fact, it
Darwin’s theory of evolution. To apply a GA, an initial is recognised that GAs cannot provide a certificate of global
population of individuals is created at random, with each optimality. For each run of the algorithm, the best solution found
individual representing a solution to the problem. Next, pairs of by the GA will be displayed; the term ‘optimal solution’ is
individuals are combined by means of a crossover operation to avoided here. The feasibility of the solutions found by the GA
produce offspring for the next generation. This leads to an will be discussed as well.
intense search in one area of the solution space. In order to
explore more areas of the solution space, a mutation process is Chromosome representation
also used to randomly modify some of the individuals of each
generation. The process is repeated until the stopping criterion The chromosome representation (string) is intended to represent
the solution vector of the problem, and also handle as many
is reached, and the best solution found is retained. The
constraints of the problem as possible. In this case, the objective
probability of crossover for an individual is proportional to its
is to find the fixed pressure and location combination of fans in
fitness; the fitness of an individual is a measure of the quality of
the available installation locations of the mine, in order to
the solution. The use of a genetic algorithm leads to an provide the required airflow with minimum cost or minimum
improvement of the individuals in successive generations, deviations from requirements. The fitness functions will be
according to Bodenhofer (2007) and Goldberg (1989). The explained for each analysis. As the current practice in the mining
pseudocode of the GA used in this paper is given as follows. industry is to account for the maximum requirement over a
Begin certain amount of time, the idea is to test these two different
Compute initial random population V (size n) fitness functions and demonstrate their potential to generate
Evaluate individuals of V with 3D-CANVENT ventilation savings compared to manual solutions.
solver The proposed representation consists of the pressure value for
each fan location. Pressures are defined as a discretised range of
Compute fitness of individuals of V values at which the fan can operate. The range for booster and
While Termination criterion is not reached do main fans are the same: ranging from 0 to 4000 pa with
Select pairs of individuals from V for crossover discretised steps of 100 pa. Possible fan locations result from the
E ACUÑA et al
MF
WF2
WF5
WF3
F2
WF4
F1
WF1
FIG 1 - Simplified schematic of the ventilation network.
physical restrictions of the mine, and are an input to the problem.

The chromosome representation, as shown in Table 1, will TABLE 1
specify a fixed pressure for each fan location that will be used by Chromosome (string).
the ventilation solver (3D-CANVENT) to generate the airflows
of the network. 1500 200 0 400
Fan fixed Fan fixed … Fan fixed Fan fixed
Population size and number of generations pressure at pressure at pressure at pressure at
position 1 position 2 position p - 1 position p
The population size, n, is the number of individuals maintained
through each generation of the GA; these individuals provide the
basis for generating improved solutions. This parameter will Equation 8). The problem with this approach is that the airflow
requirements are not included in the objective function, nor in
impact the performance of the GA, but it has to be determined
the chromosome representation, so the GA will not be able to
experimentally.
find a feasible solution. To correct this problem, a large preset
penalty term is added to the fitness when the solution is
Constraint handling infeasible. This forces the search to concentrate on the feasible
Several different constraints must be considered in the context of region.
mine ventilation, as outlined by Lowndes and Yang (2004). In
Pj ⋅ Qj
this study, the only constraint considered is to supply at least the
required airflow to the different work faces. It is understood that ∑ 1000
+ Penalty _ value (8)
j ∈FanLocations
quantity is not enough; quality is also necessary. As the intention
of this study is to evaluate the performance of the different
fitness functions, the only constraint that will be considered is where:
the airflow requirements. Quality is handled through the use of Pj is the pressure of the fan in branch j
an initial oriented network without recirculating loops and a final
Qj is the airflow in branch j
check on the directions of the flows according to the initial
network. FanLocations are the branches where the fans can be located
and Penalty_value is a large preset value.
Fitness functions
Airflow deviation fitness function
Airpower fitness function The second approach is to minimise the deviation of the airflow
A first approach to minimise the energy consumption in a mine supplied from the airflow requirement, as it is generally assumed
ventilation network is to minimise the airpower that is required that if there is less deviation from the airflow required in
to supply the airflow requirement through the main ventilation different work faces, the energy costs to supply that airflow will
systems. It is generally assumed that if there is less airpower also be reduced. In that case, the fitness function can be given as
utilised to provide the airflow required in different work faces, the sum of the square of the differences between the airflow
the amount of airflow provided will also be reduced, adjusted to requirement and the airflow provided by the fixed pressure
the demand. In that case, the fitness function can be given as the combination (see Equation 9). The problem with this approach is
sum of the airpower required by the fixed pressure fans (see the symmetry of the fitness function with respect to the airflow
AN APPLICATION OF GENETIC ALGORITHMS TO THE MAIN VENTILATION SYSTEM OPTIMISATION
requirements. As an example, if the airflow requirement is 30 m3/s population sizes ranging from 50 to 300 with a step of 50.
and the airflow supplied is 20 m3/s or 40 m3/s, then the value of Resulting values and solving times are displayed in Table 3.
the fitness function will be the same, and the GA will not be able
to differentiate between a feasible solution and an infeasible one.
To correct this problem, a large preset penalty term is assigned TABLE 3
when the solution is infeasible. This forces the search to Solving time and airpower for airpower function
concentrate on the feasible region. (hours:minutes:seconds).
AP Gen Gen
∑ (Qr j − Qs j )2 + Penalty _ value (9) Pop size 50 100 150 50 100 150
j ∈FanLocations 50 0:00:24 0:00:45 0:01:06 905.93 837.68 1002.20
100 0:00:49 0:01:16 0:01:50 919.29 767.04 805.15
where:
150 0:01:12 0:01:59 0:02:50 815.97 767.04 861.16
Q is the airflow requirement in branch j 200 0:01:43 0:02:49 0:04:05 849.24 849.24 849.24
Qs is the airflow supplied in branch 250 0:02:02 0:03:24 0:04:49 880.33 815.97 805.15
AirflowLocations are the branches where the airflow 300 0:02:30 0:04:12 0:05:41 805.15 849.24 805.15
requirements are fixed by planned activities and Penalty_value
is a large preset value.
For all 18 runs, the GA was able to find feasible solutions for
the ventilation network where the airflow requirements were
Selection, crossover and mutation operators delivered at least as required. In 15 out of the 18 runs, the GA
was able to improve the manual optimised solution, and the
The GA used is the one described in the algorithm values are presented with shaded background. The best solution
implementation section, where the selection method is roulette in terms of the airpower was 767.04 kW, and compared to the
wheel and the crossover operator is one point crossover, initial optimised solution that generated an improvement of 13 per
described by Goldberg (1989). The crossover probability is cent. The best value was found in two runs out of the 18.
60 per cent and the mutation probability is ten per cent.
Airflow deviation fitness function
Baseline For the airflow deviation fitness function approach, the same 18
runs were generated as for the airpower fitness function
Hardcastle (1983) had previously determined an initial manual
approach. Solving times and results are displayed in Table 4.
optimised solution for the ventilation network according to the
operation in the field. This allows two things: first, a check of
whether good feasible solutions are generated and second, a TABLE 4
comparison of the performance of the GA against a manual Solving time and airpower for airflow deviation function
optimised solution. (hours:minutes:seconds).
The results of the original manually optimised solution are
Dev Gen Gen
presented in Table 2. Three potential fan locations are allowed in
the mine: one main fan (MF) and two booster fan (F1 and F2). A Pop size 50 100 150 50 100 150
fourth fixed pressure booster fan represents the natural 50 0:00:23 0:00:53 0:01:19 892.49 1002.20 905.93
ventilation pressure (NVP), but it is not shown on the network 100 0:00:57 0:01:32 0:02:08 880.33 815.97 861.17
diagram. The total combined fan power determined from this
initial manual solution was 881.43 kW. 150 0:01:20 0:02:20 0:03:05 861.17 933.02 861.17
200 0:02:05 0:03:19 0:04:27 1002.20 919.29 986.75
RESULTS 250 0:02:03 0:03:50 0:05:34 933.02 933.02 861.17
300 0:02:39 0:04:12 0:05:38 861.17 971.88 767.04
Airpower fitness function
Again, for all 18 runs the GA was able to find feasible
Based on the parameters and the chromosome representation solutions for the ventilation network. For this case, in only eight
described in the previous section, the GA was implemented. runs out of the 18 runs the GA was able to improve the manual
Several runs were generated to test its ability to find feasible optimised solution, and the values are presented with shaded
improved solutions in terms of the airpower consumed, in order background. The best solution in terms of the airflow deviation
to deliver the airflow requirements to the work faces. 18 runs had a value of 0.79, and the airpower associated with that
were generated with three different generation numbers solution was 767.04 kW. The best value was found in only one
(iterations): 50, 100 and 150, and with a varying number of run out of the 18.
TABLE 2
The initial manual optimised solution.
Fan No Branch From To Fixed pressure Flow Fan

1 3 2 202 2.388 192.96 MF
2 16 72 71 2.094 56.32 F1
3 25 90 91 3.996 69.9 F2
4 6 3 2 0.131 178.52 NVP
kW 881.43
E ACUÑA et al
was improved. In terms of the solving time, both fitness

functions are comparable as expected. Further work is required
1050 to determine the best set of parameters for each network and
1000
what is the impact of larger networks with a larger number of
fans in the solving time.
950
Airpower
900
ACKNOWLEDGEMENTS
850
This research was partially supported by MITACS
800 ACCELERATE, Ontario, Canada. The authors would like to
750 thank Gary Li of Natural Resources Canada and Robert Maynard
0.8 0.9 1.0 1.1 1.2 1.3 of MIRARCO for their support integrating 3D-CANVENT with
Airflow deviation the algorithm. The genetic algorithm was developed using the
GAlib genetic algorithm package, written by Matthew Wall at
the Massachusetts Institute of Technology.
FIG 2 - Airflow deviation versus airpower.
REFERENCES
Acuña, E, Hardcastle, S, Maynard, R, Fava, L, Hall, S and Dunn, P, 2009.
Comparison between both approaches The application of genetic algorithms to multiple period ventilation
systems for multilevel mine operations, in Proceedings Orebody
Modelling and Strategic Mine Planning (ed: R Dimitrakopoulos),
Values and feasibility of the solutions pp 265-270 (The Australasian Institute of Mining and Metallurgy:
Melbourne).
Both approaches were able to generate feasible solutions for the
ventilation network with the population size and generation Bodenhofer, U, 2007. Genetic algorithms: Theory and applications
[online], Fuzzy Logic Laboratorium Linz-Hagenberg. Available
number implemented. However, 83 per cent of solutions from: <http://www.flll.uni-linz.ac.at/teaching/Ga/GA-Notes.pdf>.
generated with the airpower fitness function were better than the
Calizaya, F, McPherson, M J and Mousset-Jones, P, 1987. An algorithm
initial manual optimised solution. The solutions generated with for selecting the optimum combination of main and booster fans in
the airflow deviation approach were better than the manually underground mines, in Proceedings Third US Mine Ventilation
optimised solution only 44 per cent of the time. Although both Symposium, pp 408-417.
approaches converged to the same ‘best’ optimised solution Goldberg, D E, 1989. Genetic Algorithms in Search, Optimization and
(767.04 kW), the results obtained do not show correlation Machine Learning (Addison-Wesley, Don Mills: Ontario).
between airflow deviation and airpower. Figure 2 presents the Hardcastle, S G, 1983. Analysis of airflow patterns and pollutant
airflow deviation versus the airpower for the airflow deviation concentrations in mine ventilation recirculation systems, in thesis
fitness function results. It was initially expected that a higher submitted to the University of Nottingham for the degree of doctor of
airflow deviation will have higher airpower consumption, but philosophy, volumes 1 and 2.
near the ‘best’ solution found the behaviour observed was just the Hardcastle, S G and Kocsis, C K, 2007. The ventilation Challenge – A
opposite except only by the ‘best’ solution found. The reason for Canadian perspective of maintaining a good working environment in
these results is that the fine adjustment to airflow requirements in deep mines, in Challenges in Deep and High Stress Mining,
pp 519-525.
the work faces of the network around the ‘best’ solution found
generates more expensive solutions airpower wise, but because Hardcastle, S G, Kocsis, C K, Bissiri, Y and Baiden, G, 2007. Continued
development of a process simulation approach to determine airflow
the fitness function just seeks to reduce airflow deviation, it requirements in the optimal design of a ventilation system, in
cannot recognise whether a reduced deviation has a positive Challenges in Deep and High Stress Mining, pp 537-544.
impact on airpower at the same time. Kocsis, C K, Hall, R and Hardcastle, S G, 2003. The integration of mine
simulation and ventilation simulation to develop a ‘life-cycle’ mine
Solving time ventilation system, in Application of Computers and Operation
Research in the Mineral Industries, pp 223-229 (Southern African
As expected, the solving time is similar in both approaches; one is Institute of Mining and Metallurgy).
not clearly faster than the other. The solving time for the airpower Lowndes, I S, Fogarty, T and Yang, Z Y, 2005. The application of
fitness function ranged from 24 seconds up to five minutes and genetics algorithms to optimize the performance of a mine
41 seconds, with a total time for the 18 runs of 43 minutes and ventilation network: The influence of coding method and population
26 seconds. In the case of the airflow deviation fitness function size, in Soft Computing, volume 9, pp 493-506.
the solving time ranged from 23 seconds up to five minutes and Lowndes, I S and Yang, Z Y, 2004. The application of GA optimization
38 seconds, with a total time for the 18 runs of 47 minutes and method to the design of practical ventilation systems for multi-level
46 seconds. metal mine operation, Transactions of the Institutions of Mining and
Metallurgy, Mining Technology, 113(1):A43-A58.
McPherson, M J, 1993. Subsurface Ventilation and Environmental
CONCLUSIONS Engineering (Chapman & Hall: London).
Petrov, N N and Popov, N A, 2004. Ways of improving economy and
Both the airpower and the airflow deviation fitness functions were
reliability of the mine ventilation, in Journal of Mining Science,
able to demonstrate their capacity to find improved feasible 40(5)531-536.
solutions for the mine ventilation network in terms of the airpower Wu, X S and Topuz, E, 1998. Analysis of mine ventilation systems using
and the airflow required for the different work faces, and to operations research methods, in International Transactions in
achieve the same ‘best’ result with an improvement of 13 per cent Operational Research, 5(4)245-254.
compared to the baseline. In terms of solution quality, both Xing, W and Topuz, E, 1988. Comparison of methods for determination
approaches are comparable, but the airpower fitness function of booster fan locations in underground mines, in Proceedings Fourth
outperformed the airflow deviation fitness function in terms of US Mine Ventilation Symposium, pp 355-362.
the number of times that the initial manual optimised solution
HOME
Applying Sequence Optimisation Technology to Longwall Coal

Mining
L Rocchi1, P Carter2 and P Stone3
ABSTRACT Run-of-mine (ROM) and washed coal product transportation

BHP Billiton’s Illawarra Coal (IC) operates several longwall coal
logistics are complex. ROM coal from Dendrobium is transported
extraction systems in the Bulli and Wongawilli coal seams contained via rail to Port Kembla, with CPP rejects transported by road for
within the Southern Coalfields of the Sydney Basin. IC has applied a disposal at the West Cliff Colliery. Product coal for export is then
proprietary Mixed Integer Linear Programming (MILP) mine planning transported by road to the PKCT. ROM coal from the Appin
optimisation tool called Blasor to rapidly evaluate the jointly optimal Colliery is transported a short distance by road to the West Cliff
underground development strategy and mining sequence for a number of CPP, with wash plant products subsequently transported by road to
longwall mining domains. For each development scenario, the Blasor the PKCT or BlueScope Steel. Defining longwall domains within
optimiser maximises the discounted operating cash flow (DCF) as the the relevant coal seam is also a complex matter. As longwall
objective function, subject to mining, processing and transportation
capacity constraints and product blend constraints. The paper outlines how mining results in varying degrees of subsidence, and the
Blasor, a planning tool developed specifically for open pit strategic Wollongong escarpment is a heavily populated area, definition of
planning, has been applied to the underground longwall domain sequence Coal Extraction Zones (CEZ) consistent with efficient longwall
optimisation problem. It demonstrates the value of such a tool for operations and minimal public disturbance requires a considerable
optimising strategic decisions through the evaluation of a set of carefully mine planning effort. Typically, further data must also be gathered
considered scenarios, with each scenario describing a feasible underground about the coal resource ahead of mining, which when coupled
development and transport strategy. In this application, Blasor plays the with the various planning approvals, can require significant
role of a fair valuation tool for each major scenario, in that it discovers the
amounts of time. Underground transport options for coal produced
most valuable extraction schedule for each development scenario, thus
establishing a sound basis for comparing the economic merits of by a longwall are also complicated, particularly for the Appin/
competing, and often structurally quite different, scenarios. West Cliff Colliery (see Figure 1) where several domains could be
transported to the surface by one of the hoisting systems, each of
which has its own cost structure and capacity constraints. The
INTRODUCTION Dendrobium Colliery is less complex, with a single coal hoisting
BHP Billitons Illawarra Coal mining operations are focused on the system (see Figure 2).
mining of two seams – the Bulli seam and the Wongawill seam, of Coal qualities vary by domain within each colliery. The ranges
the Southern Coalfields of the Sydney Basin, centred around the across all blocks in all domains are presented in Table 1. On a
coal export facilities at Port Kembla in New South Wales, block-by-block comparison, the West Cliff domain offers the
Australia. Illawarra Coal (IC) typically produces a high quality highest coking coal yields (the more valuable product), whilst
metallurgical coal and lesser quantities of steaming coal at annual Dendrobium tends to offer the lowest. Vitrinite, ash and specific
rates of about 8 Mt. Coal is produced from three longwall mining energy values also vary quiet widely between collieries. On a
systems at Appin, West Cliff and the recently commissioned
weighted average basis, the Appin domains offer the highest
Dendrobium Colliery. Two Coal Preparation Plants (CPP) are
supplied with coal from these longwalls, one located at West Cliff, coking coal yields. Tonnage splits (see Figure 3) also contribute
the other located near to BlueScope Steel’s Port Kembla blast to the complexity of the optimisation problem – what is the
furnaces. IC has a contract to supply 2.5 - 3.0 Mt per annum to optimum mining sequence that maximises the value of these coal
BlueScope, with the balance being delivered to the Port Kembla resources, given product prices, the costs of mining, washing and
Coal Terminal (PKCT) for transport to other Australian and transporting the coal, capacity constraints of the various
overseas customers. Resource modelling and mine planning components and the quality constraints for a blended product? A
activities are conducted using Mincom MineScape software and problem this complex is ideally suited to the application of
Runge XPAC software. The XPAC scheduler requires the user to Mixed Integer Linear Programming (MILP) technology on
determine a ‘strategic’ path through the longwall domains, where which the Blasor software is based.
these comprise a number of longwall blocks. Longwall blocks are
typically 300 m wide and 1 - 3 km in length. XPAC then defines
the detailed extraction sequence which follows this strategic
BLASOR METHODOLOGY
direction by using a myopic search of the different legal short-term Stone et al (2007) have discussed the development of the Blasor
extraction options and choosing the one which best meets the software. Essential inputs include a block model file, definition of
objective and relaxed constraints. XPAC usually has a difficult an operational model to describe material movements and
time in meeting hard blend constraints, but it can apply penalties definition of an economic model from which discounted cash
as a method of forcing compliance. flows are derived. In a novel application, the block model file can
be presented to Blasor as a virtual pit which excludes the
surrounding waste material otherwise required for a standard pit
1. MAusIMM, Manager Resource Strategy, Illawarra Coal, BHP optimisation. Some simplifications of the orthodox open pit
Billiton Limited, Old Port Road, Wollongong NSW 2505.
mining constraints are required particularly in relation to enforcing
mining continuity in a longwall domain once opened, otherwise
2. MAusIMM, Mining Engineering Manager, Resource and Business the mathematical model developed within Blasor provides a very
Optimisation, BHP Billiton Limited, 152 - 158 St Georges Terrace, realistic representation of the optimisation problem and is a global
Perth WA 6000. Email: [email protected] optimisation (ie objective function of maximising discounted cash
3. Manager Optimisation R&D, Resource and Business Optimisation, flow for the life-of-mine, subject to various constraints), as
BHP Billiton Limited, Richard Berry Building, Melbourne compared to the XPAC heuristic methodology. At the core of the
University, Melbourne Vic 3000. Blasor optimisation engine is a commercially available mixed
Email: [email protected] integer programming solver, CPLEX, a very powerful and flexible
L ROCCHI, P CARTER and P STONE
Appin 7E
Appin 7W
West Clif f 5
Appin Drif t
Appinn8
Appi 8
North Cli ff
Wes t Cliff
Appin 3 Winder & CPP
FIG 1 - Longwall domains for Appin/Westcliff collieries.
FIG 2 - Longwall domains for the Dendrobium Colliery.
APPLYING SEQUENCE OPTIMISATION TECHNOLOGY TO LONGWALL COAL MINING
TABLE 1
Coal quality variation by colliery.
Coal quality (block ranges) Units Appin Westcliff Dendrobium

Coking coal yield (min-max) % 44 - 89 70 - 87 34 - 50
Steaming coal yield (min-max) % 0.6 - 37 4 - 16 18 - 35
Vitrinite content (min-max) % 43 - 52 37 - 46 71 - 86
Ash content (min-max) % 10 - 15 10 - 14 27 - 31
Specific energy content (min-max) kcal/kg 6620 - 7030 6705 - 6990 6200 - 6490
Longwall Domain Tonnage Splits available constraint choices once an initial optimisation had been
completed. Some variables were ignored for the optimisation –
coal mined by development to establish access for the longwall
system, ventilation of the workings and the establishing of coal
37% conveying systems. It was reasoned that an optimum sequence
for the domains was the critical variable, with the selected
sequences being used to inform the XPAC data structure. This
51% specifically included mine development advance and tonnages,
thereby facilitating a more detailed calculation of cash flows.
Economic parameters and physical constraints

Blasor has an extensive range of options for applying constraints
12% which are consistent with evaluating open pit problems. Many of
these physical constraints were not relevant to the problem at
Appin Colliery Westcliff Colliery Dendrobium Colliery
hand. The constraints which were applied are briefly described
below:
FIG 3 - Distribution of coal inventory between the collieries. • Mine life: the number of periods over which the optimisation
should proceed – these could be months, quarters or years.
tool. Its direct application to mining optimisation problems is not • Discount rate: the periodic discount rate which should be
possible, however, due to the enormous number of decision applied to the cash flows (revenues less costs) calculated to
variables. To reduce the size of the problem, Blasor uses a number arise from the mining and processing of each block.
of proprietary heuristic algorithms, effectively aggregating and • Fixed costs: a time cost specified as a time series applying to
re-scaling space and time variables. The solution provided by all domains, for example, a management overhead applicable
Blasor is provably optimal within a predefined tolerance. to the aggregate of all products for a specific time period.
• Time adjustment factor: a constant specified as a time series
THE OPTIMISATION PROBLEM which can be applied to the cost of mining or processing a
block or its products to deflate or inflate the expenditure, or
IC identified two elements to the overall problem. Relocation of
adjust for foreign exchange rates.
one longwall system in the Appin/West Cliff domains was
imminent. A second and equally important question was related • Total mining rate: an upper and lower value for the aggregate
to an optimisation across all domains to determine whether the mining rate in all domains in each period, for all periods
optimisation of the Appin/West Cliff domains would be affected. defined by the Mine Life parameter.
• Number of pits mined simultaneously: the number of open
Input models pits (domains) which can be in operation at the same time.
Data describing the essential physical characteristics of the coal • Individual pit mining rate: for each pit (ie longwall), the
reserves within each longwall block was extracted from the upper and lower limit for the mining rate.
Mincom MineScape 4 General Mine Planning (GMP) package.
Quantity and quality data (typically the parameters in Table 1)
were then manipulated to establish a data structure suitable for Products
Blasor. In essence, this data structure reflected a column of cells
with the uppermost cell representing the first block to be mined Blasor fundamentally models mass flows and their transform-
in a domain and the lowermost block, the last to be mined. This ation from the mine to the market. Any mass flow can be defined
permitted differing directions of advance within a domain to be as a product. There are two types of products:
simulated, reflecting the differing ROM coal hoisting options.
1. ‘real products’ are mass flows (such as washed coking coal)
The Appin/West Cliff Colliery complex comprised six that generate revenue and thus have a non-zero price
domains for a total of 91 blocks, averaging about 15 blocks for assigned to them, as well as relevant capacity and blend
each domain (range 4 - 17). The Dendrobium Colliery comprised constraints; and
an additional two domains for a total of 42 blocks. Each domain
was presented to Blasor as a distinct virtual open pit, devoid of 2. ‘pseudo products’, which are zero priced arbitrarily defined
any surrounding overburden. The mining sequence within a mass flows (such as ROM coal or a contaminant of washed
domain is implicitly an open pit direction of advance – ie coal) with any combination of capacity and blend
downward. Constraints to force Blasor to maintain continuity constraints.
within a domain once mining has commenced are not a natural For any real or pseudo product, an upper and lower limit to the
feature of the software – this required some manipulation of the mass flow in each time period for all periods defined in the Mine
L ROCCHI, P CARTER and P STONE
Life parameter can be specified. Products could also be regarded of coal hoisting possibilities as binary decisions allowing
as contaminant masses in a real product: selection of the optimum transport path. These were not
• Blend constraints: for any product (real or pseudo), a encodeable to the Blasor mathematical model. Product blend
specification of the upper and lower limits proportion of the constraints were not applied for this evaluation on the underlying
product as a constituent of the real product, described as a assumption that all coal was saleable in a blend with coal from
time dependent variable. the Dendrobium Colliery.
The optimisation was not able to be configured to reliably
• Processing rate: for each processing plant, the upper and enforce mining continuity within a domain once commenced,
lower limits to the periodic rate of coal processing, for all
requiring an initial appraisal and culling ahead of applying an
periods defined by the Mine Life parameter.
interactive process to specifically enforce mining continuity. In
• Mine out constraints: for any selected pit (ie longwall practice, a domain could be opened for several years, then Blasor
domain), a specification as to which other domains must be would switch to another domain and complete mining there
completed before commencing mining in the selected before returning to complete the project in the first domain.
domain. Results of the initial optimisation for the six domains is
presented in Table 3. Hoisting Alternative B offered a relatively
Economic parameters high discounted cash flow (DCF), but for a significantly
discounted ROM coal inventory compared to Hoisting
• Mining costs: the total expenditure associated with the Alternatives A and D. Hoisting Alternative C also demonstrated
mining of each block, calculated as the extension of the mass
a low utilisation (57 per cent) of the Blasor inventory. Both of
of the block by the relevant unit costs per unit mass,
adjustable by a Time Adjustment Factor where so required. these results needed to be considered in the context of
contractual coal sales to BlueScope Steel over a 25 year period.
• Transport costs: the total expenditure associated with Consequently, Hoisting Alternatives A and D were carried
transporting mining products from the longwall to the CPP, forward with an adjustment of the individual Pit Mining Rate and
measured by extending the mass of the block by the relevant Mine Out constraints to enforce mining continuity in each
unit costs per unit mass. domain after mining commenced. Small changes in the key
• Processing costs: the total expenditure associated with performance indicators emerged, however Hoisting Alternative D
processing the entire tonnage of a block in the CPP, plus the remained as the best solution.
cost of transporting the products derived from the block and
delivering them to the PKCT, adjusted by a time adjustment
factor. Optimisation of the Appin/West Cliff/Dendrobium
domain sequence
• Product prices and selling costs: a time series specification
of the prices expected to be received for each product (ie Evaluation of alternative coal hoisting solutions was not necessary
coking coal steaming coal), together with the related selling for the Dendrobium Colliery. Consequently, the optimisation
or transaction costs. This is also specified as a time series. process was able to focus on the effect of blend constraints in the
production of the Illawarra Blend. The Illawarra Blend is a
Outputs product specification for coking coal, nominating ranges for ash
content, vitrinite content and various physical characteristics such
A wide variety of outputs, graphical and tabular, are possible as fluidity and reflectance. It is a natural focus point for customers,
with Blasor. These outputs allow the user to identify when each so the sequencing of mining domains to maximise the amount of
block in each domain is mined and where it is sent, together with coal mined that satisfies the Illawarra blend remains a critical
the related periodic cash flows (discounted and undiscounted) business endeavor.
and periodic mass flows. As with the previous evaluation, an initial optimisation was
completed without blend constraints, applying Hoisting
Optimisation of the Appin/West Cliff domain Alternative D for Appin/West Cliff and without enforced mining
sequence continuity in a domain. Mining continuity and blend constraints
were progressively introduced (see Table 4). It became clear after
The need to understand the next move for the West Cliff longwall the blend constraints were applied that one domain was not being
equipment within the Appin/West Cliff domains necessitated an mined. A progressive relaxation of the blend constraints revealed
initial focus on these domains. The additional complexity of two that the constraint for the ash content of the blended steaming coal
coal hoisting solutions for two of the six domains required the product was constraining the mining sequence. Enforcing mining
development of a matrix (see Table 2) describing the entire range continuity in the absence of any product blend constraints had a
TABLE 2
Coal hoisting alternatives matrix for the Appin/Westcliff domains.
Hoisting alternative Domains
appin3 appin7e appin7w appin8 northcliff westcliff5
A wc_hoist Yes Yes Yes Yes
ap_hoist Yes Yes
B wc_hoist Yes Yes Yes
ap_hoist Yes Yes Yes
C wc_hoist Yes Yes
ap_hoist Yes Yes Yes Yes
D wc_hoist Yes Yes Yes
ap_hoist Yes Yes Yes
APPLYING SEQUENCE OPTIMISATION TECHNOLOGY TO LONGWALL COAL MINING
TABLE 3
Optimisation results for Appin/Westcliff domains coal hoisting alternatives.
Hoisting alternative
KPI A B C D
Cumulative cash flow Relative to Hoisting Alternative C 126% 113% 100% 130%
Cumulative discounted cash flow Relative to Hoisting Alternative C 100% 106% 100% 108%
Mine life Years 33 24 20 32
ROM coal mined % Blasor inventory 95% 72% 57% 97%
Domains mined Number 6 5 4 6
TABLE 4
Optimisation results for Appin/Westcliff/Dendrobium domains applying coal hoisting Alternative D.
KPI Case 1 Case 2 Case 3 Case 4

Mining continuity Mining continuity Mining continuity Mining continuity
not enforced enforced enforced enforced
blend constraints blend constraints blend constraints blend constraints
not applied not applied applied relaxed
Cumulative cash flow Relative to Case 1 100% 99% 87% 89%
Cumulative discounted Relative to Case 1 100% 99% 92% 94%
cash flow
Mine life Years 33 33 28 29
ROM coal mined % Blasor inventory 100% 99% 86% 88%
Proportion of ROM % Mined 100% 100% 98% 99%
washed
Domains mined Number 8 8 7 8
small impact relative to the unconstrained results. Enforcing the both coal hoisting, and production within blend constraints. The
product blend constraints had a very significant effect on value of the mining sequences derived by this methodology was
utilisation of the coal resource, which was partially alleviated by demonstrably higher compared to mining sequences derived from
relaxing the steaming coal ash content constraint. The DCF impact the heuristic solution methodology using XPAC software.
of these blend constraints was of the order of several hundred
million dollars, compared to the unconstrained case.
ACKNOWLEDGEMENTS
The reasonably significant reduction in the proportion of coal
mined from the inventory presented to Blasor was also The authors acknowledge the consent of the management of BHP
significant. A number of blocks were mined and assigned as Billiton and Illawarra Coal for approval to publish this paper.
waste, ie not presented to the CPP, because this returned a higher Development of Blasor technology by BHP Billiton Exploration
value solution under the enforced mining continuity requirement. and Mining Technologies is also acknowledged.
This suggested that the confidence levels in coal quality data for
these blocks should be examined and appropriate actions
implemented. REFERENCES
Stone, P, Froyland, G, Menabde, M, Law, B, Pasyar, R and Monkhouse,
CONCLUSIONS P, 2007. Blended iron ore mine planning optimisation at Yandi,
Western Australia, in Orebody Modelling and Strategic Mine
The application of MILP technology has provided an optimised Planning, second edition (ed: R Dimitrakopoulos), pp 117-120 (The
solution (ie highest value) for two problems: both optimising Australasian Institute of Mining and Metallurgy: Melbourne).
HOME
Combining Optimisation and Simulation to Model a Supply Chain

from Pit to Port
P Bodon1, C Fricke2, T Sandeman3 and C Stanford4
ABSTRACT Zeigler, Praehofer and Kim (2000) describe a method for

simulating a system using a discrete event system specification.
An export supply chain, beginning with the extraction of ore from a pit
and ending with the loading of this ore onto vessels at a port, is a key These models are constructed by considering each physical item
component of many mining operations. These supply chains are (train, car dumper, reclaimer, ship, etc) as a discrete entity, with
comprised of a number of complex subsystems such as mining, ore its own uniquely defined set of properties or attributes (speed,
processing, transportation, stockyard management and vessel loading. material type, reliability, carrying capacity, etc). These entities
Typically, the operation and performance of each of these subsystems is act out the operational activities that make up the processes being
analysed in isolation, with little consideration of their interaction with modelled. They consume discrete periods of time for each
upstream and downstream subsystems. In reality, stochastic and dynamic activity, and incur delays that can be logically induced (eg bin
influences that affect one of these subsystems will have flow on effects empty, no rake, etc). They also use stochastic methods to
for all other subsystems in the supply chain. Hence, evaluation of the generate randomly induced delays (eg breakdown, failures, etc),
performance of the total integrated system needs to capture the
all of which are dependent on the data and operational rules that
interaction of these subsystems. Discrete Event Simulation (DES) has
proved to be a powerful tool in modelling supply chains, capturing the are defined for that particular process or piece of equipment.
system dynamics and interactions, and evaluating the overall performance This combination of logical and random events is designed to
of the integrated system. reflect the most likely operational environment. Each system
The primary objective of mining export supply chains is typically to within a DES model has individual operating rules and
maximise production capacity, ie tonnes of ore loaded onto vessels at the parameters which need to be accurately defined. In the context of
port. In some mining operations, the extracted ore is blended into a a mining operation, an export supply chain involves the
variety of products with differing characteristics before being exported. movement of ore from pit to port, via any number of subsystems.
This can be the case for ores such as coal, iron and manganese. In these In many mining supply chains, the operational rules regarding
operations, an additional objective, in the form of achieving a the movement of ore are simply defined (eg lump material goes
predetermined quality of material on the vessels, is equally important as a to a lump stockpile, fines material goes to a fines stockpile). The
measure of system performance as production capacity. The objective of lack of product diversification in these instances means that there
delivering a certain quality of product often conflicts directly with the
are little or no blending requirements throughout the supply
objective of maximising production capacity, resulting in an increased
level of complexity within the supply chain. In these supply chains, the chain. These simple operational rules are able to be incorporated
decision-making process of planning the movement and blending of ore into DES models of the mining supply chain relatively easily,
through the system is paramount to the overall system performance. allowing the DES model to provide a realistic representation of
Capturing this complex planning process in a DES modelling language is the export supply chain as a whole. However, in some mining
possible, but proves to be a very difficult and time-consuming task. Since supply chains, the process of moving ore from pit to port is
planning problems are often modelled and solved using an optimisation significantly more complicated. This is particularly the case
framework, an alternative approach is to decouple the decision-making when the ore is blended into a variety of products with differing
process from the simulation model, develop a stand alone optimisation characteristics before being exported, which can be the case for
model for it, and then integrate the two to create a holistic model of the ores such as coal, iron and manganese. For operations such as
supply chain. This paper describes the approach taken and presents a case
these, day-to-day movements of ore are typically planned and
study of a successful implementation on the export supply chain of PT
Kaltim Prima Coal (KPC) in Indonesia. executed by groups of experienced individuals, who match
current mining stocks and stockpile levels with a shipping plan.
The decision process by which they do so is complex, and cannot
DISCRETE EVENT SIMULATION AND be described using a simple set of rules. This limits the ability
OPTIMISATION of a DES modelling language to precisely replicate the
decision-making process that is used in practice, and hence
Discrete event simulation (DES) modelling is the process of
provide an accurate representation of the export supply chain.
emulating real world operations in a controlled environment on a
computer. This DES provides a rational and quantitative process Optimisation modelling is ideally suited for analysing complex
for increasing understanding of the potential consequences of decision-making processes, where any number of (possibly
alternate proposals. This may range from a change in operational conflicting) objectives have been identified as being desirable,
philosophies through to the commissioning of new infrastructure. though subject to constraints such as system capacity, operational
Hence DES modelling is a useful tool for both long-term limitations and time. One of the most powerful features of an
strategic decision-making and short-term planning and optimisation model is its ability to consider hundreds of
operational decisions. thousands of possibilities and determine the optimal decision in a
very short period of time. In the mining industry, optimisation
modelling has been widely applied in long-term mine planning,
1. Senior Consultant, TSG Consulting, Level 11, 350 Collins Street, particularly production scheduling problems and ultimate pit
Melbourne Vic 3000. Email: [email protected] design. It is also possible to apply optimisation modelling to the
problem of planning the movement and blending of ore through a
2. AAusIMM, Consultant, TSG Consulting, Level 11, 350 Collins
Street, Melbourne Vic 3000. Email: [email protected]
complex export supply chain, such as those described above.
This optimisation component can then be integrated into a DES
3. Consultant, TSG Consulting, Level 11, 350 Collins Street, model of the entire supply chain, enabling a holistic model of the
Melbourne Vic 3000. Email: [email protected] system to be developed. There are a number of advantages to
4. Manager Coal Technology, PT Kaltim Prima Coal, Sangatta, Kutai modelling a complex export supply chain in this manner.
Timur, Kalimantan Timur 75611, Indonesia. Automating the process of generating the plans and carrying
Email: [email protected] them out in the simulation reduces the need for human input, and
P BODON et al
aids in the process of knowledge capture and retention. In plan, subject to real life conditions and variability. A small
addition, a stand-alone optimisation model provides the ability to amount of intelligence is required within the simulation for
easily modify and test alternative planning strategies in isolation. dealing with unexpected occurrences such as bad weather
Optimisation models also have the ability to evaluate multiple shutting down pits, or pieces of equipment failing. At the end of
criteria (eg product quality, demurrage, amount of ROM the planning period, control is passed back to the optimisation
rehandling), as well as explore the effect of changing priorities model with an updated set of inputs for the next planning period.
on each of these objectives. This process is then repeated. A diagrammatic representation of
Recently, the authors developed an automated optimisation the interaction between the simulation and optimisation models
planning engine (APE) for use in DES models of complex export is presented in Figure 1.
supply chains. The APE plans the movement and blending of ore The aim of the APE is to plan movements of ore from pit to
through the system and interacts with the DES model, which ship via intermediate subsystems such as processing plants,
attempts to enact this plan under realistic conditions, hence transportation systems (rail networks or conveying systems) and
providing an accurate representation of the system dynamics of stockyards. A feature of export supply chains in the mining
the export supply chain. The design of the APE is described industry is the inclusion of buffers (stockpiles and queues)
below. between these subsystems to mitigate the impact of subsystem
performance variability on overall system performance. In the
DEVELOPMENT AND IMPLEMENTATION OF THE case of a multi-pit, multi-product blended ore mining operation,
AUTOMATED OPTIMISATION PLANNING ENGINE intermediate stockpiles are used for blending the ore into
products to be shipped, as well as for buffering purposes. It
Generally, a DES model of a supply chain will consider the follows that inputs to the APE are short-term mine and shipping
performance of the system over a one year time period, using a plans and the current levels in the intermediate stockpiles. The
mine plan and shipping plan for one year as inputs. The APE is APE will then determine the manner in which material is to be
used to plan material movements on a more frequent basis, such moved through the system via the intermediate buffer stockpiles
as fortnightly, weekly, or a number of days in advance. The time to attempt to satisfy the shipping plan. The objective is to
horizon used for the planning process is an important factor in maximise the throughput of material while keeping shipped
determining the complexity of the planning problem, hence the quality as close to target as possible, subject to equipment
computation time required to solve a problem instance with the availability constraints. It is formulated as a mixed integer linear
APE. Once a short-term plan is produced, it has to be translated program involving multiple time periods. Specific complicating
into ‘tasks’ to be carried out by the simulation. The DES model concepts related to the formulation used in the APE are
then attempts to carry out these tasks as close as possible to the discussed below.
FIG 1 - Process of interaction between the simulation and optimisation models.
COMBINING OPTIMISATION AND SIMULATION TO MODEL A SUPPLY CHAIN FROM PIT TO PORT
Non-linear constraints problem in two (planning from mine to an intermediate buffer,

then taking this solution and planning from the intermediate buffer
A direct formulation of a planning problem involving blending to the vessels). While this resulted in a small decrease in overall
will typically feature a number of non-linear constraints. This was solution time, the loss of the ability to effectively plan for ships
the case in the original formulation proposed for use in the APE.
earlier in the supply chain operation resulted in a decrease in the
A period of experimentation in the use of non-linear solvers was
quality of products loaded to vessels. Therefore, this approach to
undertaken with little success, as solution times tended to be
the planning problem was abandoned.
unwieldy. To overcome this issue, a small number of simplifying
assumptions are made to enable the formulation to be linearised,
so that the solution time of the APE is reduced considerably CASE STUDY – PT KALTIM PRIMA COAL
without losing significant detail in the plans produced. SUPPLY CHAIN
PT Kaltim Prima Coal (KPC) operates a coal mine near Sangatta
Variation in stockpile quality over time in East Kalimantan, Indonesia. Coal is mined at various grades
The major complicating factor in this area is the variation in and blended through a series of intermediate stockpiles that are
stockpile quality over time. The quality of material extracted linked by a 13 km overland conveyor (OLC) before being loaded
from a stockpile in any time period is dependent on the quality as multiple products onto a ship. To determine the potential
and quantity of material added to that stockpile in the preceding consequences of increased production and alternative production
time periods. This relationship is inherently non-linear in nature. scenarios, KPC required an understanding of the interaction
Assuming a fixed quality in each stockpile throughout the between throughput and quality, and how these are impacted by
planning period under consideration enables this relationship to any changes in infrastructure and/or operating policy. The nature
be modelled linearly. In addition, this assumption leads to the of planning the KPC operation to achieve contracted coal
possibility of a further level of planning which is more tactical qualities involves multiple objectives including throughput,
than operational, ie setting the target stockpile qualities for each blending and on time delivery onto ships. An APE was
time period. constructed that incorporated these multiple objectives, as well
as the ability to interact with the DES model. The optimisation
Through loading replicates the planning activity that is regularly performed on site
to enable the DES model to operate for an extended duration
Many port operations in mining export supply chains allow for (anywhere from one week to several years). The integration of
through loading of material onto vessels. Through loading occurs the optimisation within the simulation allowed KPC to link the
when material arrives at the port and is loaded onto a vessel effects of real life uncertainties to the strategic plans being
berthed at the port immediately, effectively bypassing the buffer developed. The addition and integration of the APE within the
stockpiles in the stockyard. From a modelling perspective, this is simulation model was a complex task. In isolation, the three key
achieved by including a splitter at the port, which allows a elements of this model (quantity model, quality model and
variable amount of feed to be directed to stockpiles in the yard, planning) are well established, however the incremental addition
and the remainder to be sent directly to the ship loading of each of these elements into an integrated DES and
operation. Initial modelling of the decision of the quantity of optimisation model exponentially increases the model
material to send to each respective destination resulted in non- complexity. The insights gained from this complex modelling
linear constraints. This non-linearity was overcome by assuming system hold the potential to revolutionise the way the KPC
that through loading is always carried out when a vessel is at operation is planned and operated.
berth, and that when this occurs, a fixed proportion of the feed is
split between the through loading operation and the yard Overall benefit
stockpiles.
The integrated DES model has helped KPC in making strategic
Building stockpiles to completion before being long term decisions, short-term planning decisions and also
turned over provides the possibility of aiding the operational decisions of
creating and evaluating weekly plans.
Ideally, each port stockpile is built to completion before being
turned over and reclaimed from to load vessels. This constraint is Strategic decision-making
included in the APE by applying a penalty for turning over a
stockpile before it is full. An integer variable is included in the The primary purpose of building the integrated DES model was to
formulation to represent the state of each stockpile (stacking or help KPC understand the likely impacts of increased production,
reclaiming), and a fixed penalty is applied when a stockpile is identify bottlenecks in the system and evaluate the effect and
turned over before it is in a reclaim state. feasibility of various potential future expansions of the operation.
The ability to easily change the inputs of the model enables KPC
Solution time to quickly understand the effect of upgrades to equipment, such as
increasing crusher capacity, improving conveyor rates and
Even the most sophisticated optimisation engines generally reliability, and increasing ship loading capacity. In addition, the
require a significant amount of time to determine the optimal integrated DES model provides the flexibility to test different
solution to a complex mixed integer linear programming problem stockpiling configurations and examine the effect of reducing the
with a large number of decision variables and constraints. One of amount of through loading.
the drawbacks of the techniques used to linearise the formulation
used in the APE is that additional decision variables and Short-term planning
constraints are added to the formulation. The purpose of the APE
is to improve the decision-making processes within a DES model The ability to take a short-term marketing plan in terms of
and facilitate the generation of realistic outputs by the DES model. shipping demand for tonnes and quality, and evaluate the
Since it is used to solve many instances of the planning problem potential of the operation to supply these tonnes given a mine
throughout a single simulation run, it is imperative that the plan, has been of great benefit to KPC. The DES allows KPC to
solution time for the APE is not excessive. One approach explored see which plans are harder to achieve, as well as indicating the
in an attempt to reduce solution times was to split the planning potential bottlenecks and areas which are causing problems.
P BODON et al
Operational decision-making Objective function

The optimisation component of the integrated DES model is able The following is a list of objectives that the APE has to optimise
to operate in stand-alone mode. This provides the ability to easily against, given the constraints listed below. The importance of
modify and test alternative planning strategies in isolation. In each of these objectives is controlled by weighting multipliers in
addition, further detailed modelling of the planning process that the solver:
occurs on a day-to-day or even hour-to-hour basis could be • Maximise throughput. Primarily this means to maximise the
included in this stand-alone model. This could provide sufficient tonnage down the OLC, however it is extended to also
detail to enable the optimisation component to be used in the maximise tonnes mined and tonnes shipped.
weekly planning sessions that are held on site, where the daily
movements of ore for the next week are determined. • Minimise deviation of both crushed ore and port stockpiles
from their assigned lower and upper bounds. Quality
deviance is calculated by creating a variable which is the
The Integrated DES model difference between the desired and actual quality, then
including this variable in the objective with a negative value
The KPC Coal Chain Integrated DES model is a large, complex to penalise this deviance.
model that incorporates many features to enable it to accurately • Minimise deviation from each vessels’ target quality.
replicate the real operation. The inputs for this model include
the: key plant and equipment capability/capacity, equipment • Minimise deviation of each vessels’ loading time from its
configuration, equipment reliability (planned and unplanned arrival time.
downtime), mine plan, and shipping plan (forecast). Equipment
capacity and reliability is determined through analysis of the Constraints
existing operation. Equipment configuration is based on the
The following are the list of constraints that the APE must
current operation and may be manipulated to simulate different operate within when finding a solution that optimises the
operating conditions. Mine and shipping plans are supplied by objective function.
KPC in the form of mine log files and the current shipping
program spreadsheet. Scenarios are arranged by manipulating
these key input parameters to test the system under different Pits
configurations. To enable the estimation of maximum system • Only mine blocks that are available: blocks must be mined in
capacity and to identify and quantify system bottlenecks, it is sequence. This is achieved by limiting the number of blocks
necessary to scale the demand and supply components to ensure that are available to be planned, based on the provided mine
that the model is tested to its capacity. The supply (mine plan) is plan sequence. For example, blocks nominated to be mined
scaled by simply scaling the coal quantity of each block in the in the month of September cannot be mined in August unless
mine plan, while the demand (shipping plan) is scaled by adding all of August’s blocks have been mined. Changing the bucket
ships to, or removing them from the plan. The primary aim of the size for the blocks affects how closely the model tracks the
APE used within the integrated DES model is to determine a provided mine plan. Whilst in the short-term the order that
sequence of tasks to deliver the required coal to the waiting the blocks are mined is critical, as the model moves further
vessels. The DES model then follows the list of tasks developed into the future, so the need for strict adherence diminishes
by the APE until such time that the tasks are completed or due to uncertainties in the orebody mapping. The ore blocks
replanning is required. The following is an outline of the are ‘binned’ by month – finer resolution would require more
functionality of the APE used in the integrated model of the KPC accurate mine plans extending possibly for a number of
Coal Chain. years.
• The APE is allowed to mine blocks that are assigned to a
Planning horizon following time period in the input mine plan, but a penalty is
incurred. Essentially, blocks assigned to an upcoming period
The planning horizon used by the APE is able to be varied. would only be mined to deal with a quality issue.
Following a testing phase it was established that a planning
horizon of 21 shifts (seven days), with plans updated every nine • Precedence constraints between blocks are not explicitly
shifts (three days), provided realistic results and did not require modelled. It is assumed that precedence requirements have
significant solution times. This configuration enabled the APE to been handled in the mine plan, and excursion from the mine
provide planning based on a seven day look ahead to ensure that plan is allowed only on a quality issue.
both near term and longer term (seven day) objectives could be • Do not exceed shovel capacity in any pit and across all pits in
optimised, but also meant that the DES model was unlikely to any shift.
become significantly out of synchronisation with the plan. • Do not mine more than the tonnage of any given block.
Optimisation formulation
Crusher stockpiles
The optimisation is a general linear program as described by
Winston (1987). As such, it has an objective function which in • Do not exceed crusher rates.
this case is a maximisation, subject to a number of constraints, of • Physical configuration constraints regarding which crusher
the general form: feeds to which crusher stockpile.
Maximise ∑ x i • Do not exceed each crushed ore stockpiles’ capacity.
• Do not exceed each crushed ore stockpiles’ maximum
Subject to ∑ a i.x i <= b i
reclaim rate.
The model is formulated in Lingo using a combination of the • Aim to have each stockpiles’ quality within its nominated
techniques described by Schrage (2003). quality range at all times.
Overland conveyor KPIs. Furthermore, testing the system at a single throughput is not
sufficient to enable the successful identification of system
• Do not exceed the maximum OLC rate. bottlenecks and inefficiencies. To enable a complete under-
standing of the system performance, the combination of the
Port stockpiles following two KPIs is the focus of the integrated DES model of
• Do not exceed each port stockpiles’ capacity. the KPC Coal Chain:
• Aim to have each stockpiles’ quality within its nominated 1. Quantity: tonnes moved from mine to ship and the
quality range at all times. utilisation of the intermediate stockpiles.
• Do not exceed total port capacity. 2. Quality: the match between the customers’ contracted
shipments and the coal that is actually loaded onto their
Vessels vessels. Quality is measured by the gross calorific value
(GCV) of the coal.
• Load vessels to their stated tonnage.
• Do not load vessels prior to their arrival time. Quantity
• Do not exceed the rated capacity of the ship loader. DES modelling allows the user to quantify the amount of extra
• Aim to have each vessel within its nominated quality range. production from a proposed capital expansion, as well as
providing valuable insight into the auxiliary effects from any
actions. This is particularly valuable in systems that contain
Outputs many interacting components, such as the KPC Coal Chain.
The following data is output from the APE to the DES model. Testing the sensitivity of the system performance to varying
The DES model then uses these outputs to control (direct) the equipment rates allows the potential benefit of changing the
flow of coal. operating philosophies used in this area of the supply chain to be
determined, and also provides a means of evaluating whether
• Tonnes to crusher stockpiles: number of tonnes from each each piece of equipment is, or could potentially be, a restriction
block in the mine plan to be sent to each crusher stockpile in or bottleneck on the overall system. The following is an example
each planning period. There is the potential that a block may of the analysis that is able to be undertaken using the integrated
be split across more than one crusher stockpile.
• Crusher stockpile to port stockpile: number of tonnes to be
sent from each crusher stockpile to each port stockpile via System Performance
the OLC in each planning period. It is possible to have more Supplied Tonnes vs Demand Tonnes
Revenue and Profit vs Demand Tonnes
than one crusher stockpile feeding a port stockpile and more
than one port stockpile being fed from the same crusher
stockpile.
Shipped tonnes
• Port stockpile to vessel: number of tonnes to be sent from
US$
each port stockpile and loaded onto each vessel in each
planning period. The APE also determines which planning
period the vessels in the queue will be loaded in.
Ship plan tonnes
Example of scenario analysis Shipped Tonnes Profit Revenue
DES modelling is a complex process that has the ability to

generate a significant quantity of output results. The FIG 2 - System response to increasing demand.
interpretation of these results requires an in depth knowledge of
the model and its outputs. To best describe the performance of
the operation under varying operating scenarios requires a
statistical comparison of results and an understanding of
statistics in general. To simplify this process, a reduced number
of key performance indicators (KPIs) have been identified that
enable a simplified and more manageable understanding of the
outputs. Just as the real system has variability associated with its
performance (due to varying equipment reliability, varying times
to perform tasks and variances in the quality of ore), so too does
the DES model. For this reason, no two actual or simulated years
will ever be the same. To cope with this fact, it is necessary to
run the DES model for the same simulated period a number of
times and then calculate the mean and standard deviation of the
results for this period. Calculating the standard deviation
quantifies the effect of variability on the process, and enables the
range of results that could be expected to be produced by the
operation under similar circumstances to be gauged. System
performance is a combination of the KPIs that describe the
performance of the system in terms of its ability to load coal onto
ships in the correct quantity and quality within reasonable time
frames. Hence system performance cannot be described alone by FIG 3 - Effect of ship loader and overland conveyor rates on
any one performance indicator, but rather is a combination of shipped tonnes.
P BODON et al
DES model of the KPC Coal Chain. In this case, the aim is to CPP and TBCT Inventories
examine the effect of modifying the ship loading rate, both in Increased OLC Rate
isolation and in conjunction with the option of increasing the
OLC rate.
To establish system performance sensitivity, the model is run
at a number of throughput levels, ie using shipping plans and
Tonnes
mine plans with different levels of demand for, and supply of,
tonnes of coal. This establishes a response curve for the system.
The response curve describes the system performance as the
demands on it are increased, and provides a visual quantification
of the benefit of differing system configurations at varying levels
of throughput, as shown in Figure 2. As can be seen, increasing
ship plan tonnes will increase the quantity moved through the
Time
system, however beyond a certain tonnage, it becomes harder to
meet quality targets. This causes increased penalty payments, CPP TBCT
meaning profit eventually suffers. Figure 3 shows that increasing
ship loading rate in isolation yields very little increase in the FIG 5 - Effect of increasing overland conveyor rate on stockpile
number of tonnes shipped at the lower throughput levels. This levels.
indicates that the ship loading configuration is sufficiently
capable of meeting the shipping plan at these lower throughput
levels. As the throughput is increased, it becomes clear that GCV Error - TBCT SL Berth
increasing the ship loading rate on its own yields no significant Base Case
increase in shipped tonnes. When the increased ship loading rate
is combined with an increased OLC rate, the higher throughput
levels show measurable increases over the base case
configuration, although it would appear that the majority of the
gain is provided by the increase in OLC rate.
GCV Error
Inspection of the stockpile levels (CPP = crusher stockpiles,

TBCT = port stockpiles) with the increased ship loading rate in
Figure 4 reveals that the port stockyard is running empty,
particularly towards the end of the year. The results from
increasing ship loader rates indicate that the system is unable to
supply sufficient coal to maintain adequate inventories in the port
stockpiles. This is evidenced by inspection of the scenario with Time
OLC rate increased only (Figure 5). This chart shows that the
system was far more capable of maintaining adequate port
FIG 6 - Difference between quality of coal shipped and quality of
inventory levels, and thus able to ship more tonnes and more
coal demanded over a one year time period.
easily match quality targets.
CPP and TBCT Inventories quality of coal with the loaded quality. Each dot in Figure 6
Increased SL Rate represents a ship. The y axis indicates the error in the quality of
the material that was loaded onto each ship. Therefore a positive
value indicates the model has shipped too much of a given
quality.
Continuing with the previous example of increasing the ship
Tonnes
loading rates, Figure 7 shows the shipped qualities from the

increased ship loading rate scenario. As a result of the low levels
of port stockpiles, it has become difficult to provide each ship
with the quality of material requested. This missmatch of quality
has a direct impact on profits, either from loss of potential
revenue when providing higher quality than required, or penalties
Time
from customers when providing lower quality than requested. As
can be seen from this small example, there are a large number of
CPP TBCT
outputs that can be produced by a DES model, describing each
aspect of the simulated process.
FIG 4 - Effect of increasing ship loader rate on stockpile levels.
CONCLUSIONS
Quality
The application of a properly developed DES model provides a
The two critical elements of the KPC Coal Chain are its ability to range of significant benefits in assessing an integrated export
deliver the correct quantity and quality of coal onto ships. In the supply chain. These benefits include the ability to assess various
KPC Coal Chain operation, quality of coal is measured by gross operating practices, including maintenance options, through
calorific value (GCV). The integrated DES model incorporates quantification of performance. In addition to operating practices,
the tracking of coal quality through the system and so has the various capital expenditures can be compared to determine the
ability to measure the quality of coal loaded onto ships. To best infrastructure for a given system. By analysing all of these
quantify the effectiveness of the operation under various options, the optimal capacity of the supply chain can be
scenarios, it is necessary to compare the contracted consignment determined along with the robustness of this capacity under
GCV Error - TBCT SL Berth

complexity is added to the export supply chain. In a mining
Increased SL Rate context, examples of ores for which this may be the case include
coal, iron and manganese. The decision-making process of
planning the movement and blending of ore through the supply
chain is paramount to the overall system performance for
operations such as these. Capturing this complex planning
process in a DES modelling language is possible, but proves to
GCV Error
be a very difficult and time-consuming task. Since planning

problems are often modelled and solved using an optimisation
framework, an alternative approach is to decouple the
decision-making process from the simulation model, develop a
stand-alone optimisation model for it, then integrate the two to
create a holistic model of the supply chain.
The case study of a successful implementation on the export
Time
supply chain of PT Kaltim Prima Coal in Indonesia shows the
benefits of taking this approach to modelling for project evaluation
FIG 7 - Effect of increasing ship loader rate on quality of and strategic mine planning purposes.
coal shipped.
REFERENCES
uncertainty. The ability of DES to investigate outcomes over
many situations makes it ideal for risk analysis. Finally, Schrage, Linus, 2003. Optimisation Modelling with Lingo, pp 197-320.
quantification removes the ‘gut feel’ approach and replaces it Winston, W L, 1987. Operations Research: Applications and Algorithms
with ‘what if’ fact based analysis. (Duxbury Press: Boston).
In the case of operations that have multiple, conflicting Zeigler, B P, Praehofer, H and Tag Gon, K, 2000. Theory of Modelling
and Simulation, second edition, pp 176-180 (Academic Press).
objectives, such as delivering a certain quality of product while
also maximising production capacity, an increased level of
HOME
Direct Net Present Value Open Pit Optimisation with Probabilistic

Models
A Richmond1
ABSTRACT To overcome the inadequacy of undiscounted pay-offs in

commonly used algorithms for open pit optimisation, it is
Traditional implementations of open pit optimisation algorithms are
designed simply to find a set of nested open pit limits that maximise the
proposed to embed a scheduling heuristic within an open pit
undiscounted financial pay-off for a series of commodity prices using a optimisation algorithm. This may be seen as an alternative avenue
single ‘estimated’ orebody model. Then, the maximum Net Present Value to that taken by mixed integer programming approaches (eg
(NPV) open pit limit is derived by considering alternate (usually only best Caccetta and Hill, 2003; Ramazan, 2007; Stone et al, 2007,
and worst-case) mining schedules for each open pit limit. Divorcing the Menabde et al, 2007a) that may become numerically demanding
open pit limit delineation from the NPV calculation in this two-step in the case of large deposits. As a consequence, uncertain and
approach does not guarantee that an optimal NPV open pit solution will be time-dependant variables such as commodity prices can also be
found. A new open pit optimisation algorithm that considers the mining incorporated stochastically into the optimisation process. This
schedule is proposed. As a consequence, it can also account explicitly for permits strategic options for project timing and staging to be
commodity price cycles and uncertainty that can be modelled by stochastic assessed as discrete optimisation problems and compared
simulation techniques. This state-of-the-art algorithm integrates Monte quantitatively and is more advanced than other recent approaches
Carlo-based simulation and heuristic optimisation techniques into a global (Monkhouse and Yates, 2007; Dimitrakopoulos and Abdel Sabour,
system that directly provides NPV optimal pit outlines. This new approach 2007). It is also proposed to consider multiple conditional
to open pit optimisation is demonstrated for a large copper deposit using simulations in the optimisation process such that the mining and
multiple orebody models.
financial implications related to small-scale grade variations are
honoured (Menabde et al, 2007b; Ramazan and Dimitrakopoulos,
INTRODUCTION 2007; Leite and Dimitrakopoulos, 2007; Godoy and
Dimitrakopoulos, 2004; Ravenscroft, 1992). By considering
Several open pit optimisation techniques such the Lerchs- discounted block pay-offs, stochastic models of commodity prices
Grossman algorithm (Lerchs and Grossman 1965), network flow and short-scale grade variations a more accurate discounted
pay-off matrix (revenue block model) is generated, which in turn
(Johnson, 1968), pseudoflow network models (Hochbaum and
will yield an open pit limit that will be closer to the true optimum.
Chan, 2000) and others, involve a 3D grid of regular blocks that
is converted a priori into a pay-off matrix by considering a 3D
block model of mineral grades and economic and mining NPV CALCULATIONS WITH UNCERTAIN
parameters. These algorithms rely on the block pay-offs VARIABLES
averaging linearly, as is the case when undiscounted block
pay-offs are considered. However, the Net Present Value (NPV) Calculation of the NPV for a given open pit limit relies on
of the block pay-offs is a non-linear function of the undiscounted estimates of numerous parameters, including (but not restricted
block pay-offs that depends explicitly on the discount to be to) the mineral grades, extraction sequence and timing, mineral
applied to the individual blocks, which in turn depends on the recovery, prevailing commodity price and capital and operating
block mining schedule. To overcome the issue of discounting costs. All of these parameters are uncertain and should be
block pay-offs, traditional implementations of open pit modelled stochastically. For example, mineral grade values by
optimisation algorithms are designed simply to find a set of geostatistical simulations, operating costs with growth functions
nested open pit limits that maximise the undiscounted financial and commodity prices using long-term mean reverting models
pay-off for a series of constant commodity prices using a single that account for periodicity. Consequently, the cumulative
‘estimated’ orebody model. Then, the maximum NPV open pit distribution of total financial pay-offs for an open pit limit can be
limit is derived by considering alternate (usually only best and derived from the combination of a series of stochastic models of
worst-case) mining schedules for each open pit limit. This mineral grades, costs, prices, recoveries, etc.
two-step approach to finding the maximum NPV open pit limit
raises three significant issues: Given L potential NPV outcomes for a block (related to L
realisations of grade values, commodity prices, etc), we can
1. divorcing the open pit limit delineation from the NPV calculate the NPV for any realisation l:
calculation does not guarantee that an optimal (maximum)
NPV open pit solution will be found; B
NPVl = ∑ d l ( b j )i j (1)
2. NPV calculations are based on a constant commodity price j=1
that fails to consider its time-dependant and uncertain
nature; and
and the expected NPV for L realisations:
3. the single ‘estimated’ orebody model is invariably
smoothed, thus it fails to consider short-scale grade 1⎧L ⎫
variations. NPVL = ⎨∑ NPVl ⎬ (2)
L ⎩l=1 ⎭
Consequently, the block model does not accurately reflect the
grade and tonnage of ore that will be extracted and processed
during mining. where:
B is the number of blocks under consideration
1. MAusIMM, Principal Geostatistician, Golder Associates, 611 dl(bj) is the discounted value for block bj for the lth realisation
Coronation Drive, Toowong Qld 4066.
Email: [email protected] ij = 1 if bj falls within the open pit limit and 0 otherwise
A RICHMOND
The idea being to find the open pit limit that maximises NPVL. bi is a successor of bj if there exists an arc directed from bj to bi.
Additional financial goals, for example minimising downside In this paper, the set of all successors of bj will be denoted as Γ j .
risk (Richmond, 2004a) could also be considered, but are outside For example, in Figure 1, Γ 8 = {2, 3, 4}. A closure of a directed
the scope of this paper. graph, which consists of a set of blocks B, is a set of blocks
Bp⊂B such that if bj∈Bp then Γ j∈Bp. For example, in Figure 1, Bp
ACCOUNTING FOR MULTIPLE OREBODY = {1 - 5, 7 - 9, 13} is a closure of the directed graph. The value
of a closure is the sum of the pay-offs of the vertices in the
MODELS closure. As each closure defines a possible open pit limit, the
Pit optimisation algorithms found in the literature invariably closure with the maximum value defines the optimal open pit
consider an orebody block model with a single grade value for limit.
each block (or parcel). In such an approach, a simple decision For simplicity of notation, the algorithm proposed in this paper
rule is used where block bj is processed using option k if is described for a single orebody model. The undiscounted
gk≤z*(bj)<gk+1, pay-off matrix {w(b), b∈B} typically used for open pit
optimisation is calculated as:
where:
gk is the cut-off grade for processing option k (by convention w(b) = tonb(vz(b)rk-ck) (3)
g1 = 0 and k = 1 indicates waste)
z* is the estimated grade value where:
To account for grade uncertainty in open pit optimisation, tonb represents the tonnage of block b
Richmond (2004a) proposed incorporating L grade values for
v is the commodity (attribute z) value per concentration
each block. In this approach, multiple grade values zl(bj),
unit
l=1,…,L were generated by conditional simulation and a
processing option kl(bj) was determined for each realisation. rk is the proportion of the mineral recovered using
Alternatively, conditional simulation provides short-scale grade processing option k
variations that permit local ore loss and mining dilution to be
ck is the mining and processing cost for k ($/ton)
readily accounted for in open pit optimisation by (Richmond,
2004a): In practice, rk and ck commonly vary spatially and v and ck
temporally. The discounted pay-off matrix {d(b|S), b∈B},
• generating geometrically irregular dig-lines (that separate ore
conditional to a mining schedule S, that is required for NPV open
and waste) based on small-scale grade simulations with a
pit optimisation is calculated as:
floating circle algorithm, and
• assimilating the dig-lines into large-scale geometrically d(b|S) = [tonb(vtz(b)rk-ck,t)]/(1 + DR)t (4)
regular blocks by a novel re-blocking method.
This two-step approach accounts for short-scale grade where:
variation, but also provides ‘recoverable’ grade and tonnage
information for large regular blocks suitable for open pit t is the time period in which block b is scheduled for
optimisation. In other words, the simulated grade models are extraction and processing
compressed without loss of accuracy so that optimisation is vt ck,t are the prevailing commodity price and operating
computationally tractable. cost at time t
DR is the discount rate
AN NPV OPEN PIT OPTIMISATION ALGORITHM
In Equation 4, discounted pay-offs are conditional to the
For the vast majority of open pit optimisation techniques a mining schedule as alternate schedules can be derived for the
directed graph is superimposed onto the pay-off matrix to same open pit closure. It is also important to note that, cut-off
identify the blocks that constitute an optimal open pit limit. To grades and consequently the processing option k, may change in
paraphrase Dowd and Onur (1993) – each block in the grid, response to commodity price and operating cost fluctuations over
represented by a vertex, is assigned a mass equal to its net time.
expected revenue. The vertices are connected by arcs in such a
The traditional floating cone algorithm decomposes the full
way that the connections leading from a particular vertex to the
directed graph problem into a series of independent evaluations
surface define the set of vertices (blocks) that must be removed if
of individual Γ j and if the sum of the pay-offs associated with Γ j
that vertex (block) is to be mined. A simple 2D example is
is positive, then bj is added to Bp. However, a positive
shown in Figure 1. Blocks connected by an arc pointing away
undiscounted value for Γ j does not imply that the discounted
from the vertex of a block are termed successors of that block, ie
value for Γ j is positive. In other words, negatively-valued
successors bi of block bj that may be mined significantly earlier
1 2 3 4 5 in the mining schedule and receive substantially less discounting
may not be carried by a more heavily discounted
positively-valued bj. Furthermore, the modified schedule may
have shifted more profitable bj into later periods and additional
6 7 8 9 10 waste blocks into earlier periods, reducing the discounted value
of the pit. As a consequence, NPV optimisation with the FCA
must consider the directed graph problem globally rather than the
traditional independent evaluation of locally decomposed Γ j.
11 12 13 14 15 To allow for discounting, it is proposed that a Direct NPV
Floating Cone algorithm (DFC) proceeds as follows:
1. select the time for initial investment (start of construction) tI;
2. define a cone that satisfies the physical constraints of the
FIG 1 - Directed graph representing 2D vertical orebody model. desired open pit slope angles;
DIRECT NET PRESENT VALUE OPEN PIT OPTIMISATION WITH PROBABILISTIC MODELS
3. define an ordered sequence of visiting blocks [1,2,…#<B] • two processing options (ore and waste), ie K = 2;
with positive w(b), by ordering the blocks bi firstly on • 60 Mt/year mill constraint;
decreasing elevation, and then for blocks with identical
elevations on decreasing value in w(bi); • 25 realisations of copper grades by Sequential Gaussian
Simulation (SGS);
4. set the open pit closure counter n = 0, the initial open pit
closure B np to a null set of blocks, and the Net Present Value • 25 stochastic simulations of future copper prices with a two
of initial open pit closure NPV n = 0; factor Pilipovic model that was modified to account for
periodicity and cap and collar aversion (Figure 2);
5. set j = 0;
• 25 stochastic simulations of operating costs with a growth
6. set j = j + 1; model (Figure 3);
7. float the cone to bj to create a new closure B np + 1 = B np + Γ j • monthly copper recoveries randomly drawn from normal
(excluding from Γ j any block that currently belongs to B np ); distribution with mean of 80 per cent and a standard
n +1 deviation of 1%2;
8. determine the schedule S for the new closure B p ;
• a fixed annual discount rate of ten per cent; and
9. calculate the discounted pay-off matrix {d(b|S), b∈B np + 1 }
using Equation 4 and the Net Present Value of the new • initial investment timings at discrete yearly intervals for five
closure using Equation 1; years.
n+1 n
10. accept the new closure if NPV - NPV > 0, whereupon
the current closure is updated into a new optimal closure, ie 350
Average
n = n + 1 and go to step 5; and 300
Simulation
11. if j<#, the number of blocks with positive pay-offs w(b), 250
then go to step 6.
Price (c/lb)
200
The deterministic floating cone algorithm presented above is
heuristic in nature and and not be optimal. Alternate Bp can be 150
generated by varying the initial investment timing (step 1), the
ordered path (step 3) and/or the mining schedule (step 8). 100
Investment timing to satisfy corporate constraints or to take 50

advantage of cyclical commodity prices can be investigated as
mutually exclusive opportunities by varying tI, which modifies 0
0 24 48 72 96 120 144 168 192 216 240 264 288 312 336 360
implicitly the mining schedule in step 8 above. For example, Time (months)
given a schedule S commencing at t = 0, the modified schedule
t‘ = t + tI . For delayed investment, the NPV for many potential
production assets will typically be reduced unless maximum FIG 2 - Thirty year future copper price simulations with mean
production/grade happens to coincide with the peak in cyclical reversion and collar and cap aversion.
commodity prices. However, for a risk averse and capital
constrained company, the shift of the capital cost into future
years may be strategically advantageous when considered in 3.5
conjunction with other mining assets. Re-initiating the test 3.3
Average
Simulation
sequence from the top of the mineral deposit each time a
3.1
positively-valued cone is found and added to the closure is
generally regarded to estimate the heuristic maximum 2.9
Waste ($/t)
undiscounted pay-off solution (Lemieux, 1979). Computational 2.7

experimentation on the ordering of blocks in step 3 above
2.5
suggested that this also holds true for the discounted case when tI
is fixed. Note that, due to re-initiation of the test sequence it is 2.3
common for B np + 1 = B np in step 7 above. For such instances, steps 2.1

8 to 10 above are ignored. 1.9
It is well known that the floating cone algorithm may not 1.7
return the maximum undiscounted pay-off solution. However, it 0 24 48 72 96 120 144 168 192 216 240 264 288 312 336 360
is used in the algorithm presented above to generate physically Time (months)
feasible solutions. The author has not investigated whether the 17

Lerchs-Grossman and network flow algorithms could be 16
Average
Simulation
substituted for the floating cone algorithm, but the non-linearity 15
of the proposed objective function may present some difficulty. 14
The computational efficiency of the proposed algorithm is 13
enhanced significantly when a simple scheduling algorithm in 12
Ore ($/t)
step 8 above is employed. However, more complex risk-based 11
scheduling algorithms to account for multiple orebody models 10

9
and production goals (eg Godoy, 2002) could be considered.
8
7
APPLICATION TO A COPPER DEPOSIT 6
5
0 24 48 72 96 120 144 168 192 216 240 264 288 312 336 360
This section demonstrates the proposed concepts for a large
Time (months)
subvertical copper deposit. The geometry and contained copper
per level are variable, but there is no strong trend. The options
considered in this study were: FIG 3 - Thirty year waste and ore processing cost simulations.
A RICHMOND
Figure 2 shows 25 stochastic simulations of future copper closures than embed the scheduler in the pit optimisation process.
prices. The assumptions in this study were: In the example shown, the DFC approach that generated a single
pit required around the same computational time as that required
• a long-term copper price of $1.30/lb, in generating 36 nested pits by a simple FC approach.
• the present time ($2.50/lb) was near the peak of the price Figure 5 shows the distribution of potential NPVs for the set
cycle, of nested FCA pits without any investment delay. As expected,
• an average eight year copper price cycle, and the uncertainty increases with pit size with some possibility of
negative NPVs for large pit closures. If minimising downside
• $0.50/lb and $3.00/lb lower and upper aversion values. financial risk is of greater importance than maximising the
Note that, as time increases uncertainty in the simulated NPV then the financially efficient set (frontier) of open pit
copper price increases and the deviation of the average simulated limits could be determined under a stochastic framework
value to the long-term price decreases. The average copper price (Richmond, 2004a).
does not fluctuate symmetrically around the long-term copper
price due to the asymmetrical aversion limits. Figure 3 shows 25 9
stochastic simulations of waste and ore processing costs. average
simulation
8
To assess the potential improvement in NPV against the
7
traditional two-stage pit optimisation approach a base case
scenario ($1.30/lb – 80 per cent recovery, $1.90/t waste cost and 6
NPV (billion $)
$8.50/t milling cost) was run to generate a series of nested pits 5
using a FCA. The E-type (or average) of the 25 SGS realisations 4
was adopted as the single grade model as it is known to be 3
smoothed. The NPV for this series of pits using the base case
2
assumptions are shown in Figure 4 as crosses. The maximum
NPV under the base case scenario is associated with a pit closure 1
of 26 402 blocks. Note that, the capital cost, which could also be 0
modelled stochastically, was not included in this study. -1
0 20000 40000 60000 80000 100000 120000 140000
Pit Size (blocks)
5
4 FC 0 year
FIG 5 - Pit size versus NPV distribution.
FC 1 year
FC 2 year
NPV (billion $)
3 FC 3 year
FC 4 year
FC 5 year
CONCLUSIONS
2 Base case
DFC 0 year A novel method for working with discounted pay-off matrices
DFC 1 year
1 DFC 2 year during open pit optimisation was proposed. The approach used in
DFC 3 year
DFC 4 year
this study embedded a simple ore scheduler in a floating
0
DFC 5 year cone-based heuritic algorithm. It was a trivial exercise to further
consider multiple orebody models, local ore loss and mining
-1 dilution, time-dependent commodity prices and costs and
0 20000 40000 60000 80000 100000 120000 140000
Pit Size (blocks)
variable metal recoveries during optimisation. As a consequence,
alternate project development timings could be strategically
FIG 4 - Pit size versus NPV (FCA = floating cone algorithm; assessed. Traditional evaluation of a set of nested pit shells with
DFC = proposed direct NPV FCA). constant metal prices and operating costs failed to determine the
maximum NPV pit under uncertain conditions. However, provided
that sufficient pit shells were generated and evaluated with the
The NPV for the FCA nested pits were also calculated using same stochastic price and cost input as for the proposed algorithm
the simulated grades, metal prices, costs and recoveries for the there was little difference in the maximum NPV shell derived.
six annual investment timings, shown in Figure 4. Note that: Further experimentation should be undertaken to determine
• These curves vary substantially from the base case. whether this observation holds for more complex mining schedule
algorithms and geometrically irregular orebodies, as well as when
• In all instances the maximum NPV pit is significantly larger
a smoothed block model other than the E-type of the stochastic
(49 239 - 85 093 blocks) than the base case and the
maximum NPV is higher than for the base case. grade model is used to generate a series of nested closures.
This study demonstrated that uncertainty in future metal prices
• Delaying the investment from Year 3 to Year 5 results in a and operating costs cannot be adequately captured in open pit
higher NPV ($3.02 billion versus $2.88 billion). At first this optimisation by simply post-processing a series of nested pit
relationship appears counter-intuitive as costs are greater and closures with constant values. Stochastic modelling of mineral
discounting greater. However it is related to higher Cu prices grades, mineral recovery, commodity prices and capital and
in key production periods. operating costs provide an ideal platform to:
The NPV of the proposed DFC approach for the six annual • generate an optimal pit to maximise the overall project NPV
investment timings are also shown in Figure 4. Note that, considering geological and market uncertainty,
considering the mining schedule explicitly in the optimisation
process was successful in finding the maximum NPV pit in a • determine the optimum investment and project start up
single run. Whilst the improvement over the maximum NPV pit timing, and
from the two-step approach that considered the stochastic inputs • quantify the multiple aspects of uncertainty in a mine plan.
was limited (usually <0.5 per cent in NPV), there was often some
difference in the pit dimension. It is likely that these differences The example studied in this paper indicates periods of
would be reduced further if additional pit closures had been potential financial weakness that could benefit from management
generated for evaluation in the two-step approach. Computat- focus (eg forward selling strategies and placing the mine on care
ionally, it was more efficient to post process a finite series of pit and maintenance) prior to difficulties arising.
DIRECT NET PRESENT VALUE OPEN PIT OPTIMISATION WITH PROBABILISTIC MODELS
REFERENCES Monkhouse, P H L and Yeates, G, 2007. Beyond naive optimisation, in

Orebody Modelling and Strategic Mine Planning, second edition
Caccetta, L and Hill, S P, 2003. An application of branch and cut to open (ed: R Dimitrakopoulos), pp 3-8 (The Australasian Institute of
pit mine scheduling, Journal of Global Optimisation, 27:349-365. Mining and Metallurgy: Melbourne).
Dimitrakopoulos, R and Abdel Sabour, S A, 2007. Evaluating mine plans Ramazan, S, 2007. The new fundamental tree algorithm for production
under uncertainty: Can the real options make a difference?, scheduling of open pit mines, European Journal of Operations
Resources Policy, 32:116-125. Research, 177:1153-1166.
Dowd, P A and Onur, A H, 1993. Open pit optimisation – Part 1: Optimal Ramazan, S and Dimitrakopoulos, R, 2007. Stochastic optimisation of
open-pit design, Transactions of the Institutions of Mining and long-term production scheduling for open pit mines with a new
Metallurgy, Mining Technology, 102:A95-A104. integer programming formulation, in Orebody Modelling and
Godoy, M, 2002. The effective management of geological risk in Strategic Mine Planning, second edition (ed: R Dimitrakopoulos),
long-term production scheduling of open pit mines, PhD thesis, pp 310-333 (The Australasian Institute of Mining and Metallurgy:
University of Queensland, Brisbane. Melbourne).
Godoy, M C and Dimitrakopoulos, R, 2004. Managing risk and waste Ravenscroft, P J, 1992. Risk analysis for mine scheduling by conditional
mining in long-term production scheduling, SME Transactions, simulation, Transactions of the Institutions of Mining and
316:43-50. Metallurgy, Mining Technology, 101:A101-108.
Hochbaum, D S and Chan, A, 2000. Performance analysis and best Richmond, A J, 2004a. Integrating multiple simulations and mining
implementations of old and new algorithms for the open-pit mining dilution in open pit optimisation algorithms, in Proceedings Orebody
problem, Operations Research, 48:894-914. Modelling and Strategic Mine Planning (ed: R Dimitrakopoulos),
Johnson, T B, 1968. Optimum open pit mine production scheduling, PhD pp 63-68 (The Australasian Institute of Mining and Metallurgy:
thesis, University of California, Berkeley, 120 p. Melbourne).
Korobov, S, 1974. Method for determining ultimate open pit limits, Richmond, A J, 2004b. Financially efficient mining decisions
department of mineral engineering, Ecole Polytechnique, Montreal. incorporating grade uncertainty, PhD thesis, Imperial College,
London, 208 p.
for open pit mine planning: Application and risk analysis at a copper Richmond, A J and Beasley, J E, 2004. An iterative construction heuristic
deposit, Transactions of the Institutions of Mining and Metallurgy, for the ore selection problem, Journal of Heuristics, 10(2):153-167.
Mining Technology, 116:A109-A118. Stone, P, Froyland, G, Menabde, M, Law, B, Pasyar, R and Monkhouse,
Lemieux, M, 1979. Moving cone optimising algorithm in Computer P, 2007. Blended iron-ore mine planning optimisation at Yandi
methods for the 80s, 329-345 (SME of AIMMPE: New York). Western Australia, in Orebody Modelling and Strategic Mine
Planning, second edition (ed: R Dimitrakopoulos), pp 117-120 (The
Lerchs, H and Grossman, I F, 1965. Optimum design of open pit mines, Australasian Institute of Mining and Metallurgy: Melbourne).
Bulletin of Canadian Institute of Mining, 58:47-54.
Whittle, J, 1999. A decade of open pit mine planning and optimisation –
Menabde, M, Froyland, G, Stone, P and Yeates, G A, 2007. Mining The craft of turning algorithms into packages, in Proceedings APCOM
schedule optimisation for conditionally simulated orebodies, in ’99 Computer Applications in the Minerals Industries 28th
Orebody Modelling and Strategic Mine Planning, second edition International Symposium, pp 15-24 (Colorado School of Mines:
(ed: R Dimitrakopoulos), pp 292-310 (The Australasian Institute of Golden).
Menabde, M, Stone, P, Law, B and Baird, B, 2007a. Blasor – A
generalised strategic mine planning optimisation tool, 2007 SME
Annual Meeting and Exhibit.
HOME
Pushback Design of Open Pit Mines Under Geological and

Market Uncertainties
C Meagher1, S A Abdel Sabour2 and R Dimitrakopoulos3
ABSTRACT In conventional open pit planning, block values are estimated

Strategic open pit mine planning is carried out under multiple sources of
using a single point estimate of metal content (ore grade), a fixed
technical and financial risks. Given the uncertainty about ore grade, metal flat metal price and a fixed exchange rate (for example, Caccetta
prices and foreign exchange rates, assigning dollar values to mining and Giannini, 1988; Ramazan, 2007). An example of commonly
blocks at the planning time in order to optimise block destinations (mill, used formulas for estimating block value can be expressed as
stockpile or waste) is a challenge. In such an uncertain environment, there (Whittle, 1988, 1999; Hochbaum and Chen, 2000):
is a probability that the actual grade, metal prices and exchange rate at the
production time will be different from those based on which an optimum
block destinations have been chosen at the planning time. In such cases,
V = ToGRP − ToC p − TC m (1)
the mine management has the flexibility to revise the original decision
regarding the block destination that has been made at the planning time where:
and send the block to the optimum destination based on the new
information. Another issue that affects the block value and consequently V = block value, $
the optimisation process is the fact that the value of a block extracted
sooner is greater than the value of the same block extracted later. This To = tonnes of ore in the block
time value of money is usually ignored in block valuation. However, it G = grade, unit/tonne
could have a significant impact on the optimisation process.
This paper aims to quantify the value of management flexibility to R = recovery
revise or alter original decisions and integrate it into the dollar value P = unit price, $/unit
assigned to each mining block at the planning time, while addressing the
time value of money concept. In this respect, this work outlines real Cp = processing cost, $/tonne
options valuation (ROV) approach for jointly handling stochastically
described geological and market uncertainties and integrating the block
T = total amount of rock (ore and waste) in the block
destination flexibility in the process of assigning a value to mining blocks Cm = mining cost, $/tonne
at the planning time. The output of this ROV model is a time-dependent
discounted expected value for each block that captures the value of Block value estimations using current, common conventional
management flexibility rather than a single static value. A parametric procedures are based on three main implied assumptions:
minimum cut algorithm is then applied to produce single pushback
1. the ore grade or metal content of each block is known with
designs. The algorithm used intrinsically leverages the uncertainty in a set
of simulations of the orebody to produce low risk open pit pushback
certainty,
designs. The method used for optimising over multiple realisations of an 2. market variables such as metal prices and exchange rates
orebody, as opposed to single geostatistically estimated block values and are known with certainty, and
optimising over a single orebody model, leads to an increase in the
probability that the pushbacks produced for the earlier stages of the mine 3. at the project evaluation stage it is assumed that there are
meet their expected value. no possible future revisions of decisions related to optimum
block destinations.
INTRODUCTION This means that the decision whether to send a block to the
mill or to the dump has to be taken at the time of planning for all
The aim of open pit mine planning is to define optimum pit
blocks and no future revisions to these decisions are allowed. In
limits and an optimum life-of-mine (LOM) production schedule
other words, at the project evaluation stage, it is assumed that all
that maximise the pit value under some technical and operational
decisions regarding block destinations have to be made upfront.
constraints. The basic input to this process is a set of block
These observations have also been reported by Henry, Marcotte
values representing the net economic worth of each block. Based
and Samis (2004).
on the estimated block values, the optimiser selects the optimum
destination of each block so as to maximise the overall pit value The first assumption indicates that there is perfect knowledge
under some given technical constraints. A dollar value is usually about metal content of mining blocks. This clearly contradicts the
assigned to each block by estimating the revenue of recoverable fact that some, but not complete, information about metal content
metal at a given fixed metal price and subtracting applicable is provided by borehole data and other exploration techniques.
mining, processing and other costs. Therefore, the value of each These data are then interpolated and/or extrapolated to develop
mining block depends, among other factors, on its metal content, orebody models using geostatistical modelling techniques. This
metal prices and foreign exchange rates. usually results in a single, average value estimate for the metal
content of each mining block. Given the limited information and
the embedded assumption about the mineralised material between
1. COSMO – Stochastic Mine Planning Laboratory, Department of boreholes, the probability that the estimated metal contents will be
Mining and Materials Engineering, McGill University, FDA realised is small. Accordingly, there is a high probability that the
Building, 3450 University Street, Montreal QC H3A 2A7, Canada.
actual metal content will be different from that used to calculate
block values. Figure 1 shows the estimated average copper grade
2. AMEC Americas Ltd, 111 Dunsmuir Street, Vancouver BC of a mining block at a copper deposit, as well as 20 equally
V6B 5W3, Canada. Email: [email protected] possible realisations. It is obvious that the actual block grade could
3. FAusIMM, Professor and Director, COSMO – Stochastic Mine be significantly higher or lower than the estimated average. Given
Planning Laboratory, Department of Mining and Materials the highly variant possible outcomes in Figure 1, expressing block
Engineering, McGill University, Montreal QC H3A 2A7, Canada. grade in a single number oversimplifies reality and could lead to
Email: [email protected] erroneous block valuations.
C MEAGHER, S A ABDEL SABOUR and R DIMITRAKOPOULOS
1.2
assume for example that both the copper price at the optimisation
time and its expected long-term level equal US$2.00/lb. Figure 2
1
shows ten paths over a period of five years. Each path represents
a possible scenario for the future copper prices. Based on
Cu Grade, %
0.8
conventional mine planning procedures, a fixed copper price of
Simulations
US$2.00 will be used to calculate block values based on which
0.6
Average optimum decisions will be made at Year 0, for each block,
whether to send it to the mill or to the dump. Consider blocks
0.4
scheduled to be extracted in Year 5 and classified as waste based
on a constant copper price of US$2.00/lb – will they be sent to
0.2
the dump in Year 5 even when the copper price follows the path
represented by Realisation-4 in Figure 2? On the other hand, for
0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
blocks that are scheduled for the mill in Year 5 at the price of
US$2.00/lb, will the operator proceed with this schedule without
Simulations
re-checking even if the actual copper price follows the path in
FIG 1 - Possible Cu grade of a mining block. Realisation-6? Obviously, what happens in reality is that mine
operators usually revise previously taken decisions with time
based on the new information. This is usually carried out by
Ignoring grade uncertainty could lead to suboptimal mine plans re-running the optimiser at a regular time intervals or when
with significant deviations from production targets. The ultimate major changes to the key variables take place. It is important
outcome has an adverse effect on the project bottom line since here to note that, standing at the early project evaluation stage,
project value was optimised on the basis of inaccurate inputs. The the conventional optimiser cannot bring such a flexibility to
significance of modelling and integrating geological uncertainty modify block destinations in the future.
into open pit mine planning has been emphasised in
Dimitrakopoulos Farrelly and Godoy (2002) and Dimitrakopoulos, 5
Martinez and Ramazan (2007). Uncertainty about ore grade was
identified as a critical source of risk affecting mining project Realization-4
viability. Godoy and Dimitrakopoulos (2004) and Leite and 4

Dimitrakopoulos (2007) show through case studies that integrating Realization-7
Cu price, US$/lb
geological uncertainty in open pit mine planning significantly

reduces the risk of deviation from production targets and could 3
result in 26 - 28 per cent increase in project value. Realization-2

Realization-5
The second implied assumption in conventional open pit mine 2

Realization-1
Expected
planning is that market variables such as metal prices and Realization-3
exchange rates are fixed, ie do not change throughout Realization-10

Realization-9
life-of-mine (LOM) and are known with certainty. Obviously, 1
Realization-8
this assumption is far from realistic – looking at the past history Realization-6
of metal and currency markets, it is not difficult to conclude that 0

the probability of metal prices and exchange rates remaining 0 1 2 3 4 5 6
unchanged is null in both the short and the long-term. Therefore, Time, years
assuming that these variables will be constant throughout a
moderate LOM of, for instance, five to ten years will most likely FIG 2 - A sample of possible copper price realisations.
result in either over- or under-valuation of mining blocks. In both
cases, deviations from scheduled capacity and suboptimised
project value are likely outcomes. It is worth noting here that Previous work of integrating uncertainty in open pit mine
when evaluating projects in countries other than the USA, it is planning has been carried out by Dowd (1997), Dimitrakopoulos,
important to model the uncertainty of both metal prices and the Farrelly and Godoy (2002), Godoy and Dimitrakopoulos (2004),
exchange rate. Modelling uncertainty of metal prices expressed Leite and Dimitrakopoulos (2007) and Dimitrakopoulos,
in the local currency and assuming that this also covers Martinez and Ramazan (2007). Their focus was on integrating
uncertainty about exchange rate might be inappropriate. This is geological uncertainty into long-term mine planning. Handling
because metal prices are governed by the global supply and both metal price and exchange rate uncertainty has been dealt
demand in the world markets, while the exchange rate is with in literature focusing on real options applications to mine
governed by economic variables specific to the two countries. project valuations, which includes, among others, Brennan and
The third assumption indicates that mining blocks throughout Schwartz (1985), Trigeorgis (1993), Moyen, Slade and Uppal
LOM will be sent to the destinations optimised at the planning (1996), Kelly (1998), Moel and Tufano (2002), Monkhouse and
time with no possible future revisions. This is obviously a logical Yeates (2007), Abdel Sabour and Poulin (2006) and Samis et al
result of the first and second assumptions about certain ore grade (2006). The focus of this real options work was on market
and market variables. In other words, if ore grade and market uncertainty – geological uncertainty was not considered. Work
variables throughout LOM were known with certainty at the that focuses on both geological and market uncertainty includes
planning time, it would be possible to optimise block Henry, Marcotte and Samis (2004); Dimitrakopoulos and Abdel
destinations and consider this a final decision. In such a case, it Sabour (2007).
would be unnecessary to review the preoptimised block
destinations in the future since no changes were made to the This paper aims to integrate geological and market uncertainty
input data based on which these destinations were optimised. as well as operating flexibility to revise block destination in open
Therefore, it could be concluded that the third assumption would pit mine planning. This will contribute to enhancing the planning
be realistic if the first and the second assumptions are realistic. process in a number of ways. First, stochastic models for market
Since, as described above, block grades and market variables are variables that capture both variability and volatility will be
highly uncertain and consequently block values are uncertain, the adopted in mining block valuation rather than using a fixed price
assumption that there is no revision to previously optimised and exchange rate as in conventional analysis. Second, the value
block destinations could be erroneous. To clarify this point, of management flexibility to change decisions regarding block
PUSHBACK DESIGN OF OPEN PIT MINES UNDER GEOLOGICAL AND MARKET UNCERTAINTIES
destinations in the future based on prevailing market conditions 50
is integrated and included in block value estimates. Third, pit 4
planning will be carried out using discounted block values that 40 3

account for the time value of money. The next section outlines
Block Value, 000$CAN

the proposed methodology, followed by a case study which 2
provides a practical application, and finally, conclusions and 30
recommended future extensions are presented.

20 1
UNCERTAINTY-BASED MINE PLANNING

APPROACH
10
Expected, static
The proposed approach consists of two optimisation stages: Expected, flexible
1. a dollar value is assigned to each mining block, considering

the different aspects of uncertainty and management
flexibility to revise block destination at the production time FIG 3 - Comparison between the conventional and the
according to the realised market variables; and uncertainty-based block valuations.
2. the second stage includes a new method for designing a
lower risk long-term mine plan based on a generalisation of value, as depicted in Figure 4. In Case 3, both ore grade and
the minimum cut algorithm for ultimate pit design. copper price uncertainties are considered while exchange rate is
The method uses the simulation and block valuation data from assumed to be certain. In this case, the expected block value is
the first stage to decide when best to extract a given block to 112 per cent higher than that in Case 1, as indicated in Figure 3
reduce risk and increase profit. and the range of block value is wider than that in Case 2, as shown
in Figure 4. Moving to Case 4, where uncertainty related to all
Block valuation variables is considered, the expected block value is 134 per cent
higher than that in the deterministic case. It is worth noting here
As described above, ore grade and market variables are highly that since management has the flexibility to revise block
uncertain and cannot be defined with a point estimate. However, destination according to real time information, minimum block
modelling uncertainty of these variables does not provide much value is limited to the mining cost, incurred to give access to
useful insight nor does it make significant differences in open pit valuable block underneath, while maximum value is unlimited and
mine optimisation unless accompanied by modelling management is directly related to the three uncertain variables.
actions and responses to uncertainty resolutions. To illustrate the
effect of integrating uncertainty and flexibility into mining block
valuation, consider the simple case provided in Table 1 for a 320
mining block of 10 000 tonnes scheduled for production after five Expected
270
years. For simplicity, assume that copper grade, copper price and Maximum
CAN$/US$ exchange rate can be one of the two possibilities Minimum
220
outlined in Table 1. Figure 3 shows valuation results for this block
Block value, 000$CAN
in four cases. In Case 1, no uncertainty is considered and the block

170
has been valued using conventional methods where ore grade and
market variables are assumed to be known with certainty. Block
120
value corresponding to this case is represented by the straight line
called ‘expected, static’ in Figure 3. In this case, there is only a 70
single value estimate for the block since all variables are assumed 1
2 3 4
to be known with certainty. In Case 2, only ore grade uncertainty 20

is considered while both copper price and exchange rate are
assumed to be known with certainty. Since uncertainty is -30
considered, flexibility of management to take action is integrated
as well. As shown in Figure 3, the expected block value in this
case is 62 per cent higher than that in Case 1. This result agrees FIG 4 - Valuation of a mining block in a hypothetical copper
conceptually with the findings of Godoy and Dimitrakopoulos deposit.
(2004) and Leite and Dimitrakopoulos (2007). Also, in this case,
possible block value is spread over a range rather than a single
The first step in block valuation is to model market uncertainty.
In this respect, a variety of stochastic models are used to describe
TABLE 1 the evolution of metal prices and exchange rates with time.
An illustrative example for a hypothetical block in a copper deposit. Among them, the mean-reversion model proposed by Schwartz
(1997) is widely used since it reflects the cyclic nature of metal
Mining cost, CAN$/t 2 and currency markets and estimating its parameters statistically
using historical data is straightforward. Schwartz’s model can be
Processing cost, CAN$/t ore 10
expressed as:
Process recovery 0.9
Treatment, CAN$/t concentrate 80 dS / S = η(μ − ln S )dt + σdz (2)
Refining, CAN$/lb copper 0.1
Concentrate grade, % 30 where:
Ore grade, % Cu 0.2 or 0.5 (Expected = 0.35) S is the spot value of market variable
Cu price, US$/lb 1.00 or 3.00 (Expected = 2.00) μ is the logarithm of the long-term equilibrium level
Exchange rate 0.7 or 1.25 (Expected = 0.9) η is the reversion speed
σ is the standard deviation algorithm. The ultimate pit is then broken up into continuous
smaller sections known as pushbacks, phases or cut-backs these
dz is an increment in a standard Wiener process sections have the property that if removed in the appropriate
More details about stochastic models can be found in Schwartz order they obey the engineering slope requirements for the pit.
(1997) and Dixit and Pindyck (2004), among others. Based on These pushbacks are often produced using a Lerchs-Grossman
the stochastic model described above, a large number of correlated ultimate pit type algorithm with the orebody model scaled by
realisations of market variables are generated at each period. some parameter. Through scaling, the ultimate pit algorithm can
produce a series of nested pits which can be used as possible
The second step is to evaluate each mining block, given the
choices for pushbacks.
generated realisations of market variables, simulated orebody
models, mining cost, processing cost and all other refining and With recent advances in simulation techniques, new methods
smelting terms. At this time, a first level optimisation process is for producing ultimate pit limits and pushbacks are needed for
carried out to define the optimum destination of each block that multiple realisations of the same deposit. Averaging the multiple
maximises block value. In this respect, ROV can be applied to realisations into a single model and using the traditional
integrate the value of management flexibility to choose block techniques does not leverage some of the upside available from
destinations at the actual extraction time so as to maximise block the simulations. The method described is an extension of the
value. A comparison is carried out between the value of sending network flow/minimum cut approach to ultimate pit design (for
each block to the dump or to the processing plant. The value of a example, Hochbaum and Chen, 2000).
block if it is sent to the dump is simply the mining cost: Given an orebody model and economic values associated with
each block and a designation as either waste or ore, the minimum
VD = C M T (3) cut algorithm for producing an ultimate pit begins by constructing
a directed graph G (for a background in graph theory and a
where: definition of term, see Diestel, 2005). The graph G consists of a
node for each block in the orebody model and two extra nodes, a
VD is the block value if it is sent to the dump source node s and a sink node t. If a block bi is designated as ore
CM is the unit mining cost an arc from the source node s to the node representing bi is present
in G, the capacity of the arc (s, bi) is equal to the economic value
T is the block tonnage
of block bi. If a block bi is designated as waste an arc (bi, t) from
The block value if it is sent to the processing plant is a function the node representing block bi to the sink t is in G, the capacity of
of metal(s) price(s), exchange rate, and metal content, in addition this arc is the absolute value of the economic value of mining
to mining, processing, smelting and refining costs such as: block bi. Slope constraint are represented as arcs from a node vi to
a node vj if block bj must be removed prior to block bi. Slope
VP = F(S, MC , R, FOREX, T,C M ,C P ,C S ,C R ) (4) constraint arc have an infinite capacity.
Figure 5 shows a small 2D example of a potential block model
where: with economic values associated with the blocks. Figure 6 shows
the graph that would be constructed from that example, not the
VP is the block value if sent to the processing plant
slope constraint arcs have infinite capacity (and no labels).
S is the metal price
MC is the metal content
R is the recovery
FOREX is the exchange rate
CM is the unit mining cost
CP is the unit processing cost
CS is the smelting cost
CR is the refining cost
The decision regarding optimum block destination at the FIG 5 - A 2D vertical cross-section of a block model.
extraction time is taken so as to maximise block value:
VB = max( VD , VP S, FOREX, MC ) (5)
It is worth noting here that the block value VB is conditional

on the simulated realisation of metal price and exchange rate at
time t and the simulated orebody model. This process generates
distribution for block value at all possible extraction times. At
this stage, no operational or technical constraints such as mill
capacity and angle of slope are considered. All of these
constraints will be taken into account in the second stage of
optimisation, explained below.
Parametric minimum cut for multiple realisations

and block valuations
Traditional methods of long-term mine planning are geared
towards a single orebody model. From the single orebody model,
ultimate pit limits are calculated using the slope constraints and a
fixed economic value per block by the Lerchs-Grossman FIG 6 - The graph G constructed from the example in Figure 5.
A cut of a directed graph is a set of edges such that after the single parameter λ. The capacities on arcs leaving the source s
removal of these edges no directed path exists between the must be non-negative non-decreasing functions of λ and the arcs
source node s and the sink node t. A minimum cut is the set of entering the sink t must be non-negative non-increasing functions
edges where the sum of capacities is as small as possible over all of λ. With these properties, a series of nested pits can be
cuts in the graph. Many efficient algorithms are known for produced for λ1, λ2,…, λk where λ1<λ2<…<λk efficiently. For a
computing the minimum cut of a graph in polynomial time and small value of λ a small pit will be produced, as λ is increased
they perform well in practice. larger and larger nested pits will be created until the ultimate pit
A minimum cut, F, in the graph G corresponds to a valid pit, no is reached.
slope constraint can be contained in the set F or else the sum of Given multiple orebody realisations, one would ideally like a
capacities would be infinite but choosing the set of edges leaving long-term mine plan that met production targets and achieved a
the source s is a smaller cut, a contradiction to the minimality of F. high net present value (NPV) over all realisations. A standard
This implies that the set of blocks corresponding to nodes not approach to produce a mine plan from multiple simulations is to
reachable from s in G - F forms a valid pit in terms of slope average the realisations into a single orebody model and produce
constraints. By the way the arc capacities were specific, the the long-term mine plan in the traditional way. The problem with
minimum cut corresponds to the pit where the sum of the this approach is that much of the information on risk due to
capacities of arcs from the source to ore nodes left outside the pit geologic local variability is lost when the realisations are
(s, bi ) plus the sum of capacities of arcs from the waste nodes averaged together. For example, if you had a block worth -$3000,
inside the pit to the sink (bj, t) is minimised. This minimises the -$3000 and $9000 in three different realisations then the average
sum of ore left outside the pit plus waste inside. Since the total value would be $1000. Similarly, a different block may be worth
sum of ore inside the whole orebody model is a constant, this is $1000 in all three realisations and would be equal to $1000 when
equivalent to maximising the ore inside the pit minus the waste averaged together. There is no way to distinguish that there is a
inside the pit. Figure 7 shows the minimum cut in our example, lower risk associated with the second block than the first.
the minimum cut depicted has value 4 + (2 + 2 + 2) = 10 and Consider a small 2D example constructing the graph associated
minimises the value ore left outside the pit plus the cost of the with each simulation and merging the source nodes s and sink
waste in the pit. nodes t, the resulting graph would look like Figure 8 (the slope
constraints have been omitted from the drawing).
Simple parametric functions on the arcs from the source will be
considered, mainly the economic value of the associated block
multiplied by λ. Since the same decision must be made in the
mine plan for a given block across all simulations, the nodes in the
different simulations should be on the same side of the minimum
cut, this can be modelled by placing bidirectional infinite arcs
between corresponding nodes. Since these bidirectional arcs have
infinite capacity they will never be in the minimum cut, so the
nodes can be merged into a single node. Arcs from s to nodes that
were merged can be replaced by a single arc with the sum of the
capacities. Arcs to t from the merged nodes can also be replaced
by a single arc with the sum of the capacities. The graph on the
left of Figure 9 depicts the merged graph from Figure 8. The graph
on the right of Figure 9 is the graph if the simulations were
averaged into a single orebody model.
The graph constructed in the fashion described contains more
information from the simulations than the graph constructed
from the averaged simulations. If one had to decide between the
FIG 7 - The minimum cut in the graph G. block b1 that had values worth $-3000, $-3000 and $9000 in
three different realisations and the block b2 worth $1000 in all
The same algorithm can be used to produce a series of nested three realisations, b2 would be chosen since the difference
pits similar to those produced by Whittle (1999). A parametric between including b1 and not including it is (9000λ) - (6000),
minimum cut algorithm replaces the capacities on the arcs while the difference between including b2 and not including it is
leaving the source and arcs entering the sink with functions of a 3000λ, for 0<λ<1, 9000λ-6000<3000λ. The goal of this
FIG 8 - Minimum cut graphs with source S and sink T merged.
Consider the following example, where we are given the

horizontal cross-section of an orebody block model’s values in
three different periods. Figure 11 shows the arcs associated with
the upper left most block. If we remove it in the first period, no
cost is added to the minimum cut. If it is removed in period 2 the
arc with value 2000 - 1800 crosses the cut and represents a loss of
a potential $200. Similarly, if we remove it in period 3 the arcs
with capacity 2000 - 1800 and 1800 - 1600 cross the cut
representing a loss of $400. If the block is left in the ground the
block all three arcs from s to the block cross the cut representing a
loss of $2000.
By converting the network described into a parametric network
by multiplying the capacity of arcs from s to a node in period i by
λi and running a parametric minimum cut algorithm choosing
appropriate values of λi for each period, one can produce a series
of nested pits over the multiple periods and choose the pits that
best meet the production schedule requirements as the pushback
design. The choice of λi's can be accomplished by a Lagrangian
relaxation approach. Where the λi's for which the pushbacks
FIG 9 - The graph on the left is constructed using the simulation closely match the desired requirements are chosen. The two
data, the graph on the right is the graph constructed if the generalisations of the minimum cut algorithm, to multiple
simulations were averaged together. realisations of the orebody model and to block evaluations based
on time of extraction, can combined to produce a set of pushbacks
with a lower risk than the traditional single orebody model
modelling is to favour blocks that have a higher probability of approach.
being ore, this approach may be too conservative in some cases.
For instance, a block could have a greater than 50 per cent
chance of being more valuable than another but also a greater CASE STUDY – APPLICATION TO A
probability of being waste, this approach will favour the block COPPER DEPOSIT
with the higher probability of being ore. An interesting area of This section provides an application of the proposed procedures
further research would be to try and find a way to control the explained above to a copper deposit. The deposit is located in a
balance between risk and reward. typical archean greenstone belt. The region consists predominantly
By modelling the graph appropriately, parametric minimum cut of mafic lavas with lesser amounts of intermediate to felsics
algorithms can be used to incorporate distinct prices based on the volcaniclastics. Rocks are moderately deformed with a prominent
period of extraction. Given a series of decreasing block prices cleavage subparallel to what is considered to be the original
p(1,i)>p(2,i)>...>p(k,i) for an ore block i and periods 1 to k, the bedding, an E-W trend with average 64° south. The deposit itself
graph G will have one node per period for each block i. Let v(j,i) is in a sequence of moderately to strongly foliated, sulfidic, mafic
be the node representing block i in period j, if j < k then arc (s, to intermediate volcanic rocks, which have been intruded by
v(j,i)) will have capacity p(j,i) - p(j + 1,i) which is positive since numerous subvolcanic felsite and feldspar porphyry and/or
p(j,i)>p(j + 1,i). Let arc (s,(v(k,i)) have capacity p(k,i). If an ore intermediate volcanic tuff, with size ranging from lapilli to
block i is removed in period j the arcs from s to v(l,i) for l<j will agglomerate, within a strongly chloritic and biotitic matrix. It can
cross the cut and the sum of the capacities on these arcs represents
be traced over a strike length of 1.5 km with a thickness varying
the loss of revenue associated with waiting until period j to remove
from a few metres to more than 75 m. Mineralisation consists of
the ore. For a waste block i and decreasing costs associated with
about ten per cent sulfides, mostly chalcopiryte, pyrite and
removing the block p(1,i)<p(2,i)<…< p(k,i) for periods 1 to k, the
pyrrhotite, occurring as disseminations, streaks and stringers
arcs from v(j,i) to the sink t will have capacities p(j, i) − p(j + 1, i) .
The arc from node v(k,i) to t will have capacity p( k, i ). If the apparently controlled by the strong rock cleavage. The geological
waste block i is removed in period j, the sum of the arcs that cross database is compounded by 185 drill holes with ten metre copper
the minimum cut associated with block i equals the cost of composites in a pseudo-regular grid of 50 m × 50 m covering an
removing block i in period j. The graph G contains infinite approximately rectangular area of 1600 × 900 m2 – the average
capacity arcs from every node to the corresponding node in the dip is 60° north. Using the geological information, one
subsequent period, ie (v(j,i), v(j+1,i)) for all i and j<k. This mineralisation domain is defined and modelled through a
ensures that a block that is removed in a period stays removed in geostatistical study. In this example, there are two sources of
subsequent periods. Infinite slope constraint arcs exist between market uncertainty, related to copper price and US$/CAN$ rate.
nodes in the same period. This market uncertainty is modelled by generating 20 000
FIG 10 - Values over three different periods.
FIG 11 - Arcs associated with the upper left most block.
correlated realisations of copper prices and exchange rates using CONCLUSIONS

the mean-reversion model in Equation 2. Uncertainty about
orebody is modelled by simulating 20 orebody models using This paper outlined a simulation-based procedure for integrating
conditional simulations. Block valuations are carried out using the both geological and market uncertainties into open pit mine
procedures explained above and the data in Table 2. Figure 12 planning. This two-level optimisation procedure combines a
shows discounted values for a sample of mining blocks using both number of uncertainty modelling, advanced financial valuation
and open pit planning algorithms. As a first level optimisation,
the conventional and the proposed uncertainty-based procedure
management flexibility to change or revise block destination was
provided that these blocks will be extracted in year 10. It is integrated into block values. The described pushback design
obvious, in this case study, that conventional block valuation algorithm can use the geological and market uncertainty to
method tends to undervalue mining blocks since it ignores produce lower risk long-term mine plans. Application to a case
flexibility to revise block destination. As shown in Figure 12, the study of a Canadian copper deposit showed significant differences
difference between the two sets of valuation ranges, in block value estimates between the conventional and the
approximately, from ten per cent to 50 per cent and it is inversely proposed procedures.
related to block value. However, it is worth noting here that in
some cases, conventional procedure might overvalue mining ACKNOWLEDGEMENTS
blocks.
The work in this paper was funded by NSERC CDR Grant
TABLE 2 335696 and BHP Billiton, and the members of the COSMO
Stochastic Mine Planning Lab – AngloGold Ashanti, Barrick,
Data of the illustrative copper mine.
BHP Billiton, De Beers, Newmont, Vale and Vale Inco.
Number of blocks 21 3001
Block tonnage, t 10 800 REFERENCES
Mining cost, CAN$/t 2 Abdel Sabour, S A and Poulin, R, 2006. Valuing real capital investments
using the least-squares Monte Carlo method, The Engineering
Processing cost, CAN$/t ore 10
Economist, 51:141-160.
Recovery 0.9 Brennan, M J and Schwartz, E S, 1985. Evaluating natural resource
Treatment, CAN$/t concentrate 80 investments, Journal of Business, 58:135-157.
Caccetta, L and Giannini, L M, 1988. An application of discrete
Refining, CAN$/lb copper 0.1
mathematics in the design of an open pit mine, Discrete Applied
Concentrate grade, % 30 Mathematics, 21:1-19.
Discount rate, % 8 Diestel, R, 2005. Graph Theory – Graduate Texts in Mathematics, third
edition, Volume 173 (Springer-Verlag: New York).
Long-term copper price, US$/lb 1.60
Dimitrakopoulos, R and Abdel Sabour, S A, 2007. Evaluating mine plans
Long-term exchange rate 0.9 under uncertainty – Can the real options make a difference?
Resources Policy, 32:116-125.
Dimitrakopoulos, R, Farrelly, C T and Godoy, M, 2002. Moving forward
160 60 from traditional optimisation – Grade uncertainty and risk effects in
140
open pit design, Transactions of the Institutions of Mining and
50
Conventional block valuations Metallurgy, Mining Technology, 111:A82-A88.
Block Value, 000$
120 Real options valuation

Difference, %
Dimitrakopoulos, R, Martinez, L S and Ramazan, S, 2007. A maximum
Difference, %
40
100
upside/minimum downside approach to the traditional optimisation
80 30
60
20 Dixit, A K and Pindyck, R S, 1994. Investment Under Uuncertainty
40 (Princeton University Press: New Jersey).
10
20 Dowd, P A, 1997. Risk in minerals projects – Analysis, perception and
0 management, Transactions of the Institutions of Mining and
1 2 3 4 5 6 7 8 9 10
Mining Blocks
Godoy, M C and Dimitrakopoulos, R, 2004. Managing risk and waste
mining in long-term production scheduling, SME Transactions,
FIG 12 - Valuation results for a sample of ten blocks. 316:43-50.
Henry, E, Marcotte, D and Samis, M, 2004. Valuing a mine as a portfolio Ramazan, S, 2007. The new fundamental tree algorithm for production
of European call options – The effect of geological uncertainty and scheduling of open pit mines, European Journal of Operations
implications for strategic planning, in Geostatistics (eds: O Research, 177:1153-1166.
Leuangthong and C V Deutsch), pp 501-510 (Banff: Canada). Samis, M, Davis, G A, Laughton, D and Poulin, R, 2006. Valuing
Hochbaum, D S and Chen, A, 2000. Performance analysis and best uncertain asset cash flows when there are no options – A real options
implementations of old and new algorithms for the open-pit mining approach, Resources Policy, 30:285-298.
problem, Operations Research, 48:894-914. Schwartz, E S, 1997. The stochastic behaviour of commodity prices –
Kelly, S, 1998. A binomial lattice approach for valuing a mining property Implications for valuation and hedging, Journal of Finance,
IPO, Quarterly Review of Economic Finance, 38:693-709. 52:923-973.
Leite, A and Dimitrakopoulos, R, 2007. A stochastic optimisation model Trigeorgis, L, 1993. The nature of option interactions and the valuation of
for open pit mine planning – Application and risk analysis at a investments with multiple real options, Journal of Financial and
copper deposit, Transactions of the Institutions of Mining and Quantitative Analysis, 28:1-20.
Metallurgy, Mining Technology, 116:A109-118. Whittle, J, 1988. Beyond optimisation in open pit design, in Proceedings
Moel, A and Tufano, P, 2002. When are real options exercised? An Canadian Conference on Computer Applications in the Mineral
empirical study of mine closings, Review of Financial Studies, Industries, pp 331-337 (Balkema: Rotterdam).
15:35-64. Whittle, J, 1999. A decade of open pit mine planning and optimisation –
Monkhouse, P H L and Yeates, G, 2007. Beyond naive optimisation, in The craft of turning algorithms into packages, in Proceedings
Orebody Modelling and Strategic Mine Planning, second edition (ed: APCOM ’99 Computer Applications in the Minerals Industries
R Dimitrakopoulos), pp 3-8 (The Australasian Institute of Mining 28th International Symposium, pp 15-24 (Colorado School of
and Metalurgy: Melbourne). Mines: Golden).
Moyen, N, Slade, M and Uppal, R, 1996. Valuing risk and flexibility – A
comparison of methods, Resources Policy, 22:63-74.
HOME
Faulted Geological Model Simulation of the Resolution Porphyry

Copper Deposit
G Verly1, K Brisebois2, W Hart3 and J Hammitt4
ABSTRACT GEOLOGY
A significant portion of the risk in a mining operation is tied to the
geologic model that is being used to constrain the resource and reserve Deposit geology
estimates. Yet little has been published on geological model simulation in
the mining industry, perhaps because the original geologic setting has often
The Resolution Cu-Mo deposit is late-Cretaceous to early-
been disturbed by metasomatism and tectonism, making it often difficult to Tertiary in age and is hosted within a buried, fault-bounded
interpret and simulate. One general method to realistically simulate the sequence of Paleozoic and Precambrian sedimentary strata,
geology consists of first establishing a priority ranking of the geological Precambrian diabase sills and Cretaceous-aged layered
features and then simulating them one at a time. This can be done by volcaniclastic and siliciclastic rocks. Host strata are faulted and
starting with the interpreted outlines or shapes and incorporating the have been intruded by porphyry bodies of late Cretaceous age.
uncertainties as specified by the geologist during the simulation process. Mineralised breccia bodies have also been identified, with some
Two steps that could be part of this methodology are considered in this being spatially related to faults and porphyry intrusions (Hammitt
paper: simulation of fault surfaces, and simulation of layered rock types and Ballantyne, 2007). The mineralised rocks are buried
within the simulated faulted blocks. The algorithm consists of a series of unconformably beneath a 1000 m to 1500 m thick sequence of
sequential Gaussian simulations that are merged together after various barren sediments and volcanic rocks of Tertiary age. Figure 1
rescaling to account for non-stationary uncertainties. A detailed description shows a plan view and two vertical sections of a portion of the
of the procedure is given. Examples of results obtained on a porphyry-style 2007 geological interpretation.
copper deposit are provided.
Fault interpretation and uncertainty assessment

INTRODUCTION
Figure 2 illustrates a common method used by geologists for the
A significant portion of the risk in a mining operation or study is identification of simple block bounding faults where folding is
tied to the geological model that is being used to provide domains assumed to be minimal, using a distinctive lithologic ‘marker’
for mineral resource and reserve estimation and/or conditional intersected in multiple drill holes (eg a formation contact).
simulation. Simulating the geology is a common procedure in the
petroleum industry but is rather the exception in the mining
Stage 1
industry. Perhaps one reason is that the original geological setting
of a mineral deposit has often been disturbed by metasomatism Distinctive lithologic marker intercepts in each drill hole are
and tectonism, making realistic simulations of the geology visually identified in section view (red symbols) and grouped
difficult to produce. Some of the methods that can be used to based on their apparent relative degree of colinearity (or
simulate different aspects of the geology are boolean modelling coplanarity).
(de Fouquet et al, 1989), indicator simulation (Alabert, 1987),
plurigaussian simulation (Armstrong et al, 2003), probability field Stage 2
with local means (Srivastava, 2005) and potential field (Chiles et
al, 2007). A 2D marker horizon is interpreted and constructed, defined by a
The case considered in this paper is the Resolution smooth line (red dashes) connecting all controlling data points for
porphyry-style Cu-Mo deposit, located in Arizona, USA. The the marker within each geometric grouping. Zones of abrupt and
deposit is deep and sparsely drilled. The geological interpretation significant changes in position (eg elevation) between different
is complex and contains faults, metamorphosed rocks of segments are suspected locations for fault displacement. Fault
sedimentary origin, intrusions, breccias and alteration assembl- planes are then interpreted and constructed within these suspect
ages. One general method to realistically simulate such complex zones based on numerous geological principles including:
geology consists of establishing a priority ranking for each of the • observed apparent sense of displacement of the marker
geological features and simulating them one at a time, starting horizon;
with the interpreted outlines or shapes and incorporating the • actual fault intercepts in one or more drill holes (‘true snap
uncertainties as specified by the geologist (Verly, Bridebois and points’);
Hart, 2008). Details on how to simulate a fault block model
together with faulted rock types are presented in this paper along • typical fault geometries observed nearby on surface or in
underground workings; and
with some results.
• fault geometries predicted by well-established structural and
tectonic field studies within a similar geologic terrane, or
1. AMEC, Suite 400, 111 Dunsmuir Street, Vancouver BC V6B 5W3, predicted by laboratory experiments on rock.
Canada. Email: [email protected]
2. AMEC, 780 Vista Boulevard, Sparks NV 89434, USA. Stage 3
The spatial uncertainties of faults are assessed purely by distance
3. Resolution Copper Mining, LLC 102 Magma Heights, Superior AZ from controlling data points. Multiple fault snap points that are
85273, USA. Email: [email protected] spatially aligned may greatly reduce the uncertainty of a given
4. MAusIMM, Geologist, Kennecott Exploration Company, Rio Tinto, fault interpretation. Simple surfaces are then constructed that
224 North 2200 West, Salt Lake City UT 84116, USA. define the limits of all permissible fault positions (blue dashed
Email: [email protected] lines).
G VERLY et al
E=206
N=1006
N=1006
E=206
High Cu Shell
FIG 1 - Plan view and vertical sections through the geological interpretation. The black dots on the plan view indicate the drill hole intercepts.
FIG 2 - Diagram illustrating the methodology for the interpretation of simple faults (Stage 2, solid blue lines) in section and plan view, using
only lithologic markers intersected in irregularly-spaced or widely-spaced drill holes (Stages 1 and 2). Stage 3 illustrates the potential
degree of spatial uncertainty (orange polygons) inherent in the resulting fault model where all possible fault positions and orientations have
equal probability (dashed blue lines).
FAULTED GEOLOGICAL MODEL SIMULATION OF THE RESOLUTION PORPHYRY COPPER DEPOSIT
Faulted layered rock type interpretation and to the plane are calculated. These kinds of manipulations are
uncertainty assessment readily available in several resource modelling packages. Note that
the offset terminology is used for the distance from the interpreted
As described in the previous section on fault interpretation, the fault surface to the fault plane. It is not the displacement of the
boundaries for the Faulted Rock Types (layered rocks) are fault as in normal geological parlance.
interpreted and constructed on sections or in three dimensions by
first identifying the controlling drill hole data points for each The fault plane is rotated to the horizontal. The coordinates of
major stratigraphic horizon, and then constructing smooth lines, the fault information (surface offsets from the plane, uncertainty
polygons, or surfaces that pass through these controlling points and snap points) are similarly rotated. The reason for the rotation
and enclose lithologically distinct domains. As with the faults, is that most 2D simulation algorithms do not handle 2D planes of
every attempt is made to honor established geological principles of any orientation. Furthermore, the post-simulation grid manip-
stratigraphy. The resultant interpreted domains for the Faulted ulations are simplified in the rotated space. After rotation, the
Rock Types are then modelled as triangulations. Based on existing fault surface offsets are vertical distances from the horizontal
information from both drilling and surface mapping, analysis of plane. Because the angle of the fault surface with the fault plane
the geological interpretation is performed far from drill hole at any location is always small, the maximum possible offsets at
intercepts – the limit of variation for the vertical position of the recorded uncertainty locations are assumed to be vertical
domain boundaries was assessed to be ±30 m. For thickness, the after rotation. The fault surface is clipped by the other faults.
limit of variation was assessed as ±20 per cent of the interpreted Different realisations will result in different clipping. To avoid
thickness. gaps between the simulated fault blocks, the simulated 2D grid
needs to be extended sufficiently beyond the extent of the rotated
FAULT MODEL SIMULATION interpreted fault surface. The rotated fault surface offsets and the
vertical uncertainties are estimated on a dense 2D grid (Step 1d,
Seven faults defining eleven fault blocks have to be simulated. The Figure 3). For the offsets, the estimation method does not matter
general procedure consists in simulating one fault at a time as
much as the ‘draped’ information is very dense. This
a 2D surface and flagging a 3D grid accordingly. The seven
information, however, was available only on the clipped fault
3D grids corresponding to the seven faults are then merged
surface and was estimated within the rotated outline of the
together using the proper priorities. Each 2D surface simulation is
surface. The estimated values had to be extended to the limit of
constrained within a variable and asymmetrical uncertainty band-
width assessed by the geologist. The simulation methodology is the 2D horizontal grid. For the uncertainty, the available points
further described below and is illustrated in Figures 3 and 4. are much farther apart and ordinary kriging was used with a
spherical model with no nugget and very long ranges. Inverse
For each of the 7 faults distance could have been used as well. No extensions were
Step 1: Pre-simulation processing
necessary because the information was available for unclipped
1a) Get original fault information: surfaces. Figure 5 shows the estimated uncertainties on one side
- Fault surface (triangulation) of one fault.
Drape the fault surface (dense grid)
- Uncertainty information (series of points with 2 max. distances from
interpretation) Simulation
- Snap points
1b) Get fault plane + offsets:
- Work out approximate fault plane The rotated fault simulation is achieved in three steps:
- Work out fault surface offsets from fault plane (draped grid)
1c) Rotate the fault information: 1. conditional simulation of standard normal scores,
- Work out rotation to bring the fault plane to horizontal
- Rotate the offset information 2. rescaling the simulated normal scores to account for the
- Rotate the uncertainty information non-stationary and asymmetric uncertainty bandwidth, and
- Rotate the snap points
1d) Grid the rotated fault information: 3. adding the offsets to the rescaled normal scores.
- 2D horizontal grids
- Grid the offsets and the uncertainties The conditional simulation is sequential Gaussian. A unit-sill
Step 2: 2D simulation Gaussian variogram model with no nugget effect and a 1000 m
2a) 2D stationary simulation, conditional to snap points isotropic range has been used to ensure a very smooth surface. The
2b) Rescale the simulation to account for variable uncertainty
2c) Add the offsets 100 rotated fault surfaces
1000 m range was chosen after some experimentation. At this
stage, 100 realisations of ‘standardised’ fault surface deviations are
Step 3: 3D simulation available. Theoretically, the 100 simulated values at a given
3a) Rotate the 3D grid to be simulated
3b) Flag the rotated grid nodes location are standard normally distributed except within the zone
- Flag as above/below the simulated surfaces of influence of a snap point. The simulated values of one
100 fault simulations coded as 0/1 3D grids
realisation should also be standard normally distributed if the
Merge the simulated 3D grids surface extent is very large with respect to the variogram range.
Merge grids with proper priorities Figure 6 shows the validation statistics for one fault surface. The
100 simulated fault grids, 7 faults / 11 fault blocks per realisation top left histogram (a) shows that the mean simulated value is very
close to zero as expected, whereas the standard deviation of a
FIG 3 - Faulted model simulation process. given realisation is on average 0.66, which is lower than 1. This is
due to three snapping points that significantly reduce the possible
Presimulation work fluctuations and also due to the extent of the simulation being not
very large with respect to the 1000 m variogram range. The
The available information for one fault consists of the fault bottom left graph (c) indicates that the variogram is reproduced by
surface, the uncertainty information and three snap points for one the simulation even though significant fluctuations are observed
fault. The fault surface is modelled as a triangulation. The from one realisation to the next (d). Alfaro (2008) notes that wild
uncertainty information is provided as a series of locations on the variogram fluctuations between realisations are to be expected
fault surface with two maximum possible envelopes, one on each because the Gaussian variogram is not microergodic. The
side of the fault. The spacing between the locations is about 500 m probability plots (b) of the realisations are not straight lines
× 500 m. In a first step, the fault triangulation surface is sampled because the variances are less than one.
on a dense grid using a draping technique. The approximate plane A local rescaling of the simulated normal scores is needed to
of the fault is identified, and the normal offsets of the fault surface account for the non-stationary uncertainty. If the uncertainty was
G VERLY et al
1) Rotate
a) Fault 4) Flag rotated 3D grid
F1 2) Create dense 2D grid
3) Simulate fault in 2D
a) simulate deviations
b) rescale deviations to account
Fault for the uncertainty
Plane c) add fault offsets
F1
Horizontal
b) 3D grid
5) Rotate back the grid
F1 F2 F1 F2
6) Merge
grids
with
proper
+ =
priorities
FIG 4 - Faulted model simulation main steps.
Rotated Fault -- Fault SE Side Uncertainty

(grey lines). Note that the offsets from the horizontal have not yet
400 600 800 1000 1200 1400 1600 1800 2000 2200 2400 2600 been added, which is why the interpreted fault trace lies at
deviation = 0. The final simulation step is adding the fault offsets
1000
1000
40 14 40 49 74 202
onto the rescaled fault fluctuations, resulting in 100 rotated
800
800
simulated fault surfaces on a dense 2D grid.

600
600
Post-simulation work
Down-Dip
Down-Dip
400
400
34 4 6 14 30 17 77 206
Two sets of coordinates of the simulated 3D grid nodes are
available:
200
200
1. original unrotated, and

0
0
51 37 46 40 104 281 2. rotated as per the fault rotation.

-200
-200
400 600 800 1000 1200 1400

Along Strike
1600 1800 2000 2200 2400 2600 The second set of coordinates is used to flag the rotated grid as
Uncertainty (m)
3D Grid / Fault Intersept "Samples" "Snap Points"
above or below the simulated rotated fault, which is equivalent to
0. 250. flagging the original grid as one or the other side of the unrotated
fault. The previous steps – presimulation work, simulation, 3D
FIG 5 - Gridded uncertainty on one side of the fault after rotation. grid flagging – are repeated for each fault. When this is
completed, seven simulated 0/1 fault indicator grids are available,
with 100 realisations each. The final simulated fault block model
symmetrical on each side of the fault, the rescaling factor at a is obtained by merging the seven individual fault simulated models
given location would be a third of the maximum distance together with the proper priorities. Figure 9 shows eight realis-
estimated for the uncertainty at that location – meaning that the ations of the simulated fault model at mid-elevation in the model.
standard deviation of the rescaled simulated normal scores is a
third of the maximum distance. This ensures that approximately
99.75 per cent of the simulated fluctuations are within ± the Results and discussion
specified maximum (± 3 standard deviations). Simulated Table 1 shows the statistics for the simulated fault block models.
fluctuations outside that interval are reset to the maximum
To preserve confidentiality of the original data, each of the
distance. The uncertainty at a given location, however, is
interpreted fault block volumes has been reset to 100. The table
asymmetrical, and two rescaling factors are used, depending on
the sign of the simulated normal scores. The distribution of the shows that the interpreted (modelled) fault block volumes are not
rescaled simulated normal scores consists of the halves of two always reproduced. Differences from -14 per cent to +10 per cent
normal distributions with same zero mean but different standard are observed. These differences are due to the asymmetric fault
deviations. The two standard deviations are one third of the uncertainties. Some simulated fault blocks have a greater chance
maximum possible distances at that location. Figure 7 shows an to become smaller than interpreted and vice versa. The table also
example of such a distribution with 10 m and 15 m ‘standard shows there can be very large differences between two realisations
deviations’, corresponding to maximum fluctuations of -30 m and for a given fault block. For example the average simulated volume
+45 m, respectively. Note that the mean of the resulting for Fault Block No 1 is 89, with minimum and maximum values
distribution is not zero anymore. of 20 and 135. The coefficient of variation (CV) is 28.3 per cent.
One hundred realisations of fluctuations around the fault Assuming a normal distribution, there is a 95 per cent probability
interpretation are available at this stage. Figure 8 shows one that the simulated Fault Block No 1 volumes are within ± 56.6 per
section with the trace of one rotated fault interpretation (red line cent of the average.
at deviation = 0), the maximum possible deviations (blue lines on The approach used to simulate the faults is considered
top and bottom) and 20 realisations of the simulated fluctuations reasonable. Indeed, the results are statistically consistent with the
Fault Simulation - Simulated NSCO Checks
0.300 99.99
A) B)
99.9
99.8
0.250
99
98
95
Cumulative Probability
0.200 90
80
100 simulations 70
60
0.150 50
Nb. of data 5,109,700 40
mean 0.060 30
20
y
c
n
u
q
re
F
stdv 0.663
0.100 10
max 4.168 5
min -3.137
2
0.050 1
100 Simul.
0.2 Mean Simul.
0.1
0.000 0.01
-3.0 -2.5 -2.0 -1.5 -1.0 -0.50 0. 0.50 1.0 1.5 2.0 2.5 3.0
-4.00 -3.00 -2.00 -1.00 0.00 1.00 2.00 3.00 4.00 5.00
NSCO
NSCO
C) D)
1.20 1.20
0.80 0.80
io
V
m
ra
g io
V
m
ra
g
0.40 0.40 Azm/Plg=0/0

Azm/Plg=0/0
Model
Mean simulation Azm/Plg=90/0
Model
Mean Simul.
0.00 0.00
0. 200. 400. 600. 800. 1000. 0. 200. 400. 600. 800. 1000.
Distance Distance
FIG 6 - Simulated normal score statistics. (A) Mean of realisation histograms; (B) realisation cumulative probability curves (grey) plus mean
of the curves (red thick line); (C) variogram model (green) plus mean of realisation variograms (red dashes); (D) same as in (C) plus ten
realisation variograms along two directions (grey full and dashed lines).
TABLE 1
Simulated and interpreted fault block volume statistics. The
interpreted (model) fault block volumes have been reset to 100.
Fault Simulation volume Model Model to

block volume simulation
Min Max Average CV
% Chg
1 20 135 89 28.3% 100 -10.8%
2 57 147 108 16.8% 100 7.8%
3 33 114 86 22.2% 100 -13.9%
4 79 133 101 11.0% 100 1.4%
5 42 201 109 28.1% 100 8.6%
6 52 156 104 22.2% 100 3.6%
7 80 124 99 9.6% 100 -0.7%
≈ 100% 8 63 171 103 20.6% 100 3.2%
9 51 135 96 19.6% 100 -4.1%
≈ 95%
10 501 176 110 29.9% 100 10.1%
11 83 120 99 8.2% 100 -0.6%
≈ 67%
input data, the simulated faults are conditioned to the available
-30m -20m -10m 0m +15m +30m +45m snap points and the fault location uncertainties as assessed by the
geologist have been reproduced. The simulated fault surfaces are
visually reasonable and have more chance to be close to the
FIG 7 - Example of asymmetrical fault fluctuation distributions. interpretation than far away. On average, the interpreted fault
Two normal distributions with zero means but different standard block volumes are generally well reproduced by the simulation.
deviations are, in effect, ‘glued together’. In this example, the Where there are discrepancies, they can be explained by the
standard deviations are ten and 15 m. Simulated values outside asymmetrical input uncertainties. Although not necessary for the
the ±3 standard deviation interval are reset to three standard Resolution deposit, different aspects of the simulation could be
deviations. modified and/or improved. With the current simulation
G VERLY et al
Simulated fault deviations from interpretation

600 800 1000 1200 1400 1600 1800 2000 2200 2400
0 50 100 150 200 250
250 200 150 100

Deviations
50
Deviations
0 - 50 -100 -150 - 200 - 250

-250 -200 -150 -100 -50
600 800 1000 1200 1400 1600 1800 2000 2200 2400
Along strike
Interpretation (0 Deviation) Uncertainty bandwidths 20 Realisations
FIG 8 - Simulated fluctuations from one fault interpretation.
11 21 31 41
51 61 71 81
FIG 9 - Eight realisations of the fault block model at mid elevation in the model. The red lines represent the original fault interpretations.
The grey lines represent the simulated fault traces.
methodology, the total number of faults before and after reconstructed. The 3D interpretation of each layer is converted to a
simulation is static, as per the geological interpretation. The two 2D interpretation of elevations and thicknesses. 2D simulations of
semi-normal distribution rule used to rescale the fault fluctuations the elevations and thicknesses are performed, conditional to drill
could be changed to generate wider or narrower fluctuations with- hole intercepts and accounting for the uncertainties specified by
in the specified uncertainty bandwidths. The range of the Gaussian the geologist. The 2D simulation results are converted back to 3D.
variogram could be revisited. The simulation methodology could The faulted blocks are stitched back together at the very end. This
be changed to generate surfaces that are more planar. simulation methodology is further described below and is
illustrated in Figure 10.
FAULTED ROCKTYPE SIMULATION
Presimulation work
In the simulation described below, faulted rock types represent
rock types that are interpreted to have been offset by faulting. The available information consists of a 3D gridded geological
These rock types consist of stacks of subhorizontal layers within a model, uncertainty information and drill hole sample locations.
background made of two rock types separated by an unconformity. The uncertainty information is the same for all layers – a 95 per
The faulted rock types are intruded by dykes and breccias that do cent confidence interval (‘two sigma’) of ± 30 m from the
not appear to be offset by faulting. The general procedure for interpreted elevation and of ± 20 per cent from the interpreted
simulating faulted rock types consists in processing one fault thickness. Note the deviations for thickness are in terms of a
block at a time. The preintrusive/prebreccia rock types are relative percentage of the interpreted (modelled) thickness.
For each fault block

Step 1 - Pre-simulation processing 3D
1a) Get the maximum simulated fault block footprint
1b) Get 3D Indicator Grid:
- Generate an indicator grid of the rock types
- Reconstruct pre-intrusive rock types
- Extend rock types from interpreted to maximum simulated fault block limit
Repeat Steps 2, 3, and 4 for each stack of layers in the fault block
Step 2 - Pre-simulation processing 2D
2a) Create 2D grids of the elevation and thickness of each layer
2b) Create 2D conditioning sample dataset from drill-hole intercepts
Step 3: Simulation 2D
3a) 2D stationary simulations of deviations, conditional to snap points.
3b) Rescale simulated deviations
3c) Incorporate rescaled deviations to interpretations.
- Adjust elevations for gaps or overlaps
Step 4: Simulation 3D
4a) Flag 3D grid from 2D simulations
Step 5: Post-simulation processing 3D
5a) Recombine the 3D grids corresponding to the stacks into one 3D grid.
Nine sets of 100 realizations (one set per fault block)
Stitch the faulted block rock types back together
For a given realization, pick the simulated rock type according to its location and
simulated fault block number at that location
100 simulated faulted rock type grids
FIG 10 - Faulted rock type simulation process.
As previously mentioned, the faulted rock types are processed 20 per cent of the interpreted thickness – the normal scores are
within one fault block at a time. In a first step, the multiplied by ten per cent and then reset to ± 30 per cent if they
preintrusive/prebreccia faulted rock types are reconstructed exceed these values. The final simulation step incorporates the
(Step 1b, Figure 10 and Figure 11). This can be done by either interpretation and simulation together. The simulated elevation at a
using the preintrusive wireframe model, or by ‘eroding away’ the location is the interpreted elevation plus the rescaled simulated
intrusives/breccias from the grid model. The reconstructed rock deviation at that location. The simulated thickness is the
types are then extended to the maximum possible fault block interpreted thickness plus the product of interpreted thickness and
limits. Again, this could be done using the original wireframe the rescaled relative simulated deviation. Elevation corrections are
model or numerically approximated on the grid model. The made to avoid gaps and overlaps between the simulated layers
maximum extent of the simulated fault block is obtained by within a same stack.
scanning the realisations of the previously simulated fault model. One hundred realisations of elevations and thicknesses are
Each reconstructed and extended stack of layers is processed in available at this stage for each layer. Figure 12 shows one section
turn. The 3D interpretation of each layer is converted to a 2D with the trace of one layer interpreted elevation and thickness
grid of elevations and thicknesses. The few drill hole intercepts – (red lines), the maximum possible deviations (blue lines) and 20
one to five per fault block – are converted to 2D snap points. realisations of the elevation and thickness (grey lines). Three
snap points were available for the elevation but only two for the
Simulation thickness because of an incomplete hole intercept. The jagged
aspect of the thickness profiles is due to the exaggerated
A series of 2D simulations are performed for each stack: resolution of the vertical grid.
• same simulation of elevation fluctuations for all layers (will
tend to prevent crossovers), and Post-simulation work
• one simulation of thickness fluctuations per layer. One hundred realisations of one stack of layers within a fault
Standard normal scores are simulated first, followed by some block are available at this stage as a series of 2D grids of layer
rescaling to account for the uncertainty. The simulations are elevations and thicknesses. For each realisation, the layer’s top
conditional sequential Gaussian using a unit-sill Gaussian and bottom elevations are calculated and used to flag one 3D
variogram model with no nugget effect and a 500 m isotropic grid. The simulation procedure is repeated for the other stacks
range to ensure smooth surfaces. The checks at this stage consist within the fault block. The 3D grids are then merged together.
in verifying that the averages of the simulated normal scores are The procedure is repeated for the other fault blocks resulting in
close to zero. The variances of the simulated normal scores are 11 individual simulated models. The final simulated faulted rock
not expected to equal one because of the conditioning and the type block model is obtained by merging the eleven models
relatively small extent of the fault blocks compared to the together. For a given realisation and a given location, the
variogram range. simulated fault block is first identified. The simulated rock type
is then picked from the rock type model that corresponds to the
Rescaling the simulated normal scores is needed to account for fault block. Figure 13 shows eight realisations of a vertical
the uncertainty. The assumption is that the simulated fluctuations section through the faulted rock type model at about mid
far from snap points are normally distributed with 95 per cent of northing in the model.
the values within ± the specified uncertainty. In other words, the
specified uncertainty is equivalent to two standard deviations of
the normal distribution. The uncertainty for the elevation is 30 m –
Results and discussion
the simulated normal scores are multiplied by 15, and reset to ± 45 Table 2 shows the statistics of the simulated faulted rock type
m if they exceed these values. The uncertainty for the thickness is model. To preserve confidentiality, each of the interpreted rock
G VERLY et al
FIG 11 - Faulted rock types presimulation steps. (A) Original interpretation; (B) maximum simulated fault block footprint identified
on left and right; (C) preintrusive/breccia reconstruction; (D) extension to fault block maximum simulated footprint; and
(E) 2D grid of elevations and thicknesses of the mid-layer (pzls).
Mid-Layer Simulated Profiles

800 1 000 1200 1400
300
300
Elevation
Elevation
250
250
Elevation
200
200
150
75 150
800 1 000 Northing 1200 1400

75
Thickness
Thickness
Thickness
50
50
25
25
800 1000 1200 1400

Nor thing
Interpreted elevation/thickness Elevation +/- 30 m/thickness +/- 20 per cent
Snap locations (within 100 m of section) 20 Realisations
FIG 12 - Simulated elevation and thickness for one layer.
additional uncertainty should probably be injected in the model. A

TABLE 2 sensitivity study of the results to the range of the variogram would
Simulated and modelled faulted rock type volume statistics. be useful, as local variations could be important as the mine get
The model rock type volumes have been reset to 100. started.
Rock Simulation volume Model Model to
type
Min Max Average CV
volume simulation CONCLUSIONS
% Chg
A general method to realistically simulate the geology is
kvs 98 102 100 0.8% 100 -0.1% suggested in this paper. The method consists of first establishing a
kqs 90 113 101 4.6% 100 1.2% priority in the geological features and of simulating them one at a
pzls 89 110 99 4.7% 100 -0.7%
time. The simulation starts with the interpreted outlines or shapes
and incorporates the uncertainties as specified by the geologist.
mesc 89 114 101 4.8% 100 1.1% Two steps that could be part of such methodology were further
qzite 91 104 99 2.5% 100 -1.5% developed:
diab 97 103 100 1.4% 100 0.2% 1. simulation of a faulted block model, and
2. simulation of layered rock types within the faulted blocks.
type volumes has been reset to 100. The table shows that on The method is relatively simple because the geological features
average the interpreted (modelled) rock type volumes are are simple once they are treated one-at-a-time. For example, the
relatively well reproduced by the simulation. Differences from faults are surfaces and the faulted rock types are layers. Both
-1.5 per cent to +1.2 per cent are observed, which is considered features were simulated by simple 2D sequential Gaussian
acceptable, especially since part of these differences is related to simulations with various rescaling to account for the uncertainties
the simulated fault blocks with their asymmetric uncertainties of specified by the geologist. Last, the method is flexible in the sense
boundary positions. The table also shows the kind of differences that it can cope with complex geological models, variable
that can be observed between two realisations. For example the uncertainties and a scarcity of data. The method is made powerful
average simulated volume for pzls (Paleozoic limestone) is 99
by its incorporation of geological interpretation and uncertainty
with minimum/maximum counts of 89 and 110 respectively. The
assessments, although these can be subjective.
coefficient of variation (CV) is 4.7 per cent. Assuming a normal
distribution, there is a 95 per cent probability that the simulated
pzls volumes are within +/- 9.5 per cent of the average. These ACKNOWLEDGEMENTS
differences are significantly less than those observed for the fault
block volumes (Table 1). This is partly due to the fact that the The authors wish to thank Resolution Copper Mining, LLC and
same rock type is found in most faulted blocks. The variability of AMEC for permission to publish this paper. They also thank Harry
the rock types within any given faulted block is more significant. Parker and Geoff Ballantyne for their useful comments and
suggestions.
The approach used to simulate the faulted rock types is
considered reasonable. Indeed, the fault blocks are simulated. The
order relations between the simulated rock types are reproduced. REFERENCES
The simulated results are statistically consistent with the input Alabert, F, 1987. Stochastic imaging of spatial distributions using hard
data. The simulated rock types are conditioned to the drill hole and soft information, Master’s thesis, Department of Applied Earth
intercepts. The fluctuations in elevation and thickness reflect the Sciences, Stanford University, p 198.
uncertainty assessed by the geological staff. Still, several aspects Alfaro, M, 2008. Microergodicity and geostatistical simulation, in
of the simulation could be improved. The method assumes that Proceedings Eighth International Geostatistics Congress:
there are enough holes to get the interpretation of rock sequence Geostatistics Santiago ’08 (eds: J M Ortiz X and Emery), 1:409-417
100 per cent right at each node. This is optimistic and some (FCFM U: Chile).
11 21 31 41
51 61 71 81
C u Sh ell
FIG 13 - Simulated faulted rock type sections for eight realisations, in the northern part of the simulated area.
G VERLY et al
Armstrong, M, Galli, A, Le Loc’h, G, Geffroy, F and Eschard, R, 2003. Paul, A H, and Manske, S L, 2005. History of exploration at the Magma
Plurigaussian Simulations in Geosciences, 160 p (Springer: Berlin). Mine, Superior, Arizona, in Proceedings Geological Society of
Chiles, J P, Aug, C, Guillen, A and Lees, T, 2007. Modelling the Nevada Symposium 2005: Window to the World, pp 629-638 (eds:
geometry of geological units and its uncertainty in 3D from structural H N Rhoden).
data: The potential-field method, in Orebody Modelling and Strategic Srivastava, R M, 2005. Probabilistic modelling of ore lens geometry: An
Mine Planning, second edition (ed: R Dimitrakopoulos), pp 313-320 alternative to deterministic wireframes, Mathematical Geology,
(The Australasian Institute of Mining and Metallurgy: Melbourne). 35(5):513-544.
De Fouquet, C, Beucher, H, Galli, A and Ravenne, C, 1989. Conditional Verly, G, Brisebois, K and Hart, W, 2008. Simulation of geological
simulation of random sets, application to an argilaceous sandstone uncertainty, resolution porphyry copper deposit, in Proceedings
reservoir, in Proceedings Third International Geostatistics Congress: Eighth International Geostatistics Congress: Geostatistics Santiago
Geostatistics Avignon ’88 (ed: M Armstrong), 2:571-530 (Kluver ’08 (eds: J M Ortiz and X Emery), 1:31-40 (FCFM U: Chile).
Academic Publishers: Boston).
Hammitt, J and Ballantyne, G, 2007. Geologic interpretation and model
of the resolution deposit, unpublished proprietary data, Resolution
Copper Mining, LLC.
HOME
Geologically Enhanced Simulation of Complex Mineral Deposits

Through High-Order Spatial Cumulants
H Mustapha1 and R Dimitrakopoulos2
ABSTRACT characterisation of non-Gaussian random variables. In multiple

point statistics, training images are used as a model for high order
Earth sciences and engineering phenomena such as geologic units, grade
content and other properties of a mineral deposit, as well as attributes of
joint distributions. However this model does not necessarily
other natural phenomena, represent complex geological systems represent the true joint distribution of the random field under
distributed in space. Their spatial distributions are currently predicted consideration. The multiple-point is a particular case of the high-
from finite measurements and second-order spatial statistical models, order moment and does not infer from a concrete statistical theory.
which are limiting, as geological systems are highly complex, While cumulants that have statistical advanced properties
non-Gaussian and exhibit non-linear patterns of spatial connectivity. compared to the moments are explored in this paper and will be
Non-linear and non-Gaussian high-order geostatistics is a new area of essential in a future work. Spatial cumulants are a new concept
research based on higher-order spatial connectivity measures and spatial and it is introduced here because cumulants completely
cumulants. Key elements of a high-order spatial stochastic modelling characterise non-Gaussian stationary and ergodic spatial random
framework are developed herein, starting with the definitions of fields, thus can provide a new consistent framework in addressing
high-order spatial statistics and, more specifically, the definition and issues mentioned above. Related works on cumulants of one-
properties of spatial cumulants, and the inference and interpretation of dimensional random function models have been developed to deal
high-order anisotropic cumulants. Spatial cumulants are shown to capture with the identification, analysis and testing of non-linear signals
the directional multiple-point periodicity and spatial architecture of (eg Mendel, 1991; Nikias and Petropulu, 1993).
geological processes. It is further shown that only a subset of all the
cumulant templates has to be computed to characterise complex spatial In the following sections, this paper first outlines basic
patterns. The second key element of high-order geostatistics is the definitions, summarises approaches to calculating anisotropic
simulation of complex mineral deposits using a nonparametric Legendre spatial cumulants, shows examples, and comments on relations
series approximation with coefficient calculated in terms of spatial between cumulants and geological process. Then, the simulation
cumulants. Examples show that the approach works very well. of mineral deposits with complex non-Gaussian and non-linear
geological patterns is outlined based on the use of spatial
INTRODUCTION cumulants in the high-dimensional space of Legendre polynomials
In earth sciences and engineering, measurements of phenomena (Lebedev, 1965). Examples from simulating a complex channel
under study represent complex non-Gaussian systems distributed system illustrate the practical aspects of high-order simulations
in space. Frequently, it is required that geo-environmental based on spatial cumulants. Conclusions and comments follow.
attributes are modelled and their spatial distributions predicted
from a limited set of measurements. Random field models and HIGH-ORDER STATISTICS OF NON-GAUSSIAN
stochastic data analysis, termed geostatistics, have long been SPATIAL RANDOM FUNCTIONS
established and used as the key approach to modelling and
predicting natural phenomena in a variety of earth sciences and Let (Ω, ℑ, P) be a probability real space. A spatial real random
engineering fields (eg David, 1977; Journel, 1989; Ripley, 1987; field Z(x) is a family of random variables {Z( x 1 ), Z( x 2 ),...} at
Cressie, 1993; Kitanidis, 1997; Hohn, 1999; Goovaerts, 1997; locations, x1,x2,..., where each random variable is defined on (Ω,
Chilès and Delfiner, 1999). Despite the considerable developments ℑ, P) and takes real values.
over the past three decades, modelling approaches are based on Assuming Z(x) is a zero-mean ergodic stationary random field,
second-order statistics, and the spatial information these contain. then the cumulants of the random field Z(x) are defined from the
Concerns articulated during the last decade suggest that current MacLaurin expansion of the cumulant generating function:
modelling frameworks are limited in their ability to account for
∼ = 1n (E{exp (jvTZ)})
ψ(v)
the spatial complexity of the natural phenomena being modelled,
that are critical to modelling and predicting spatially-distributed,
location-dependent data (eg Guardiano and Srivastava, 1993). If its moments up to order r exist, then:
Several attempts to develop new techniques dealing with spatial Mom[Z(x)Z(x + h1),...,Z(x + hr -1)] = E{Z(x)Z(x + h1)
complexity include the multiple-point approach (Strebelle, 2002;
Zhang et al, 2006), new Markov random field based approaches ... Z(x + hr -1)}
(Daly, 2005; Tjelmeland and Eidsvik, 2005), and others. These
developments replace the two-point covariance with a training The moments depend only on the vectors h1, h2,...,hr -1.
image so as to account for high-order dependencies. Although Similarly, the rth-order cumulants of Z(x) can be denoted as:
these are novel approaches, there is a need for a well-defined
spatial stochastic modelling framework capable of dealing with crz ( h1 , h2 ,..., hr − 1 ) = Cum[ Z( x ), Z( x + h1 ),..., Z( x + hr − 1 )]
the complexity of geo-environmental phenomena. The approach For example, the second-order cumulant of a non-centred
advocated herein is based on cumulants, which are combinations random function Z(x) known as the covariance is given by:
of moment statistical parameters allowing the complete
c2z ( h) = E {Z( x )Z( x + h)} − E {Z( x )}2
1. COSMO – Stochastic Mine Planning Laboratory, Department of Its third-order cumulant is given by:
Mining and Materials Engineering, McGill University, Montreal QC
H3A 2A7, Canada. Email: [email protected] c3z ( h1 , h2 ) = E {Z( x )Z( x + h1 )Z( x + h2 )}
2. FAusIMM, Professor and Director, COSMO – Stochastic Mine − E {Z( x )}E {Z( x + h1 )Z( x + h2 )}
Planning Laboratory, Department of Mining and Materials − E {Z( x )}E {Z( x + h1 )Z( x + h3 )}
Engineering, McGill University, Montreal QC H3A 2A7, Canada.
Email: [email protected] − E {Z( x )}E {Z( x + h2 )Z( x + h3 )} + 2 E {Z( x )}3
H MUSTAPHA and R DIMITRAKOPOULOS
where: Figure 1 can be easily implemented in a finite difference way. If

h3 is the difference between vectors h1 and h2 r
the orthonormal r of the grid is defined with the supporting
basis
unit vector i and j. For example, the experimental third-order
It may be computationally convenient to consider zero-mean cumulant associated with the template T shown in Figure 1 is
random functions as some of the terms vanish. In addition, note given by:
that the cumulants of orders higher than three of a zero mean
1 N N r r
random function are related to their moments of lower orders and C hT13, h = 2 ∑ ∑ Z( x )Z( x + hi )Z( x + hj ),
a combination of their moments of order two. 2 N i=1 j=1
r r
{x ; x + hi ; x + hj} ∈ T3
Calculating experimental anisotropic spatial
cumulants Then, the algorithm is summarised as follows:
r r
In this section, the definitions and implementation details of the 1. Choose a template T by defining hi and hj . The angles are
calculations of experimental cumulants from exhaustive and defined by the grid: 0.45 or 90°. Start with closest grid node
sparse data sets are described. (i+1,j), (i,j+1).
Definitions 2. Loop until all possible triplets are multiplied.

3. Average the products.
Spatial cumulants are defined in terms of distances in space.
Existing cumulant calculations assume regularly sampled data 4. Increase grid node separation and go back to step 2.
and/or a regularly sampled training data set. In general, however,
geological data are available only on irregularly spaced borehole Implementation on irregular grids
locations. Similarly to anisotropic experimental variograms, it is
possible to restrict the calculation of cumulants to a given For irregular grids, the situation is computationally more complex.
direction. For this purpose, the concept of spatial template for Computations involve angles and distances calculations, that is,
calculating cumulants is introduced here. A spatial template T is the grouping of points in terms of angles and distances. In
defined as a particular geometry of points r r in space;
r more practice, the directional neighbourhood search is equivalent as
formally, given a set of directional vectors h1 , h2 ,..., hr , supported computing several directional variograms, with given angles and
by the direction angles {θ1,θ2,...,θr} the associated spatial ranges as shown in Figure 2.
template of order (n + 1) is defined (considering a spatial
location x as a reference) as:
r r r r r r
Tr + 1 ( x , h1 , h2 ,..., hr , θ 1 , θ 2 ,..., θ r ) = {x , x + h1 , x + h2 ,..., x + hr}
For example, the r=3 order cumulant with the given template
T3 is computed from:
1 N N r r
C hT13 , h 2 = 2 ∑ ∑ Z( x )Z( x + hi )Z( x + hj ),
N i=1 j=1
r r
{x ; x + hi ; x + hj} ∈ T3
Figure 1 shows an example of a template associated to a third-

order experimental cumulant.
FIG 2 - Irregular template for third order cumulant calculation. h2

and h1 are the distances with Th1 and Th2, tolerance distances
respectively. a1 and a2 are the angles with tolerances Ta1 and Ta2,
respectively. (i,j) is the basis of the cartesian coordinate system.
The algorithm for third-order irregular experimental cumulant

(Figure 3) calculations (see Figure 2 for complementary
information) can be summarised as:
r r
• Choose a template by defining hi and hj . The angles a1 and a2
can be anything between 0 and 360°. Choose tolerance
angles Ta1 and Ta2 for each of the vectors. If angle equal to
90°, omnidirectional cumulants are calculated.
• Choose a tolerance distance Th1 and Th2 for each direction.
FIG 1 - L-shape template for cumulant calculation on regular
grid data. • Find the points inside the cone angle for each vector
(Figure 2).
Implementation on regular grids • Calculate all the possible products between points inside
dashed boxes and supporting point at the origin {Z(x), Z(x1),
In the case of regular grids, the calculations are straightforward
Z(x2)}.
for particular templates that are in the same Cartesian system as r r
the grid. In this case, the distances and angles are defined in • Classify the products in terms of both distances h1 and h2
terms of grid increments. For example, the template shown in and average in every lag.
GEOLOGICALLY ENHANCED SIMULATION OF COMPLEX MINERAL DEPOSITS
SEQUENTIAL SIMULATION WITH HIGH-ORDER The discrete Legendre polynomials also satisfy:
SPATIAL CUMULANTS k
2
∑ P ( z )P ( z )Δz = 2 m + 1 δ
m i n i mn , ∀m, n ≥ 0 (5)
Approximation of a joint probability density using i=1
Legendre series
where:
This section discusses the approximation of continuous densities
using Legendre series. A squared integrable and real piecewise z = zi - zi-1 = 2/k is a space step
smooth function f defined on D = [-1,1] can be formally written k is the number of steps
in a series of Legendre polynomials:
{zi} is a uniform discretisation of [-1,1]
∞
P ( z) δmn is the delta Dirac function
f ( z) = ∑ Lm m
Pm
(1)
m=0 To avoid numerical instability in polynomial computation, we
normalised the Legendre polynomials by utilising the square
where Pm(z) is the mth -order Legendre polynomials, with norm norm. The set of normalised Legendre polynomials is defined as:
Pm , defined as (Lebedev, 1965): 2m + 1
Pm ( z) = Pm ( z)
m m 2
1 ⎛ d⎞
Pm ( z) = m⎜ ⎟
2 m! ⎝ dz⎠
[( z 2
]
− 1)m = ∑ ai , m zi , and − 1 ≤ z ≤ 1
i=0
(2) In this case, the orthogonality condition given in Equation 4
becomes:
The Legendre polynomials Pm(z) obey the following recursive k
relation: ∑ P ( z )P ( z )Δz = δ
m i n i mn , ∀m, n ≥ 0 (6)
i=1
2m + 1 m
Pm + 1( z) = xPm ( z) − Pm ( z) (3)
m +1 m + 1 −1 The coefficients Lm in Equation 1 of the Legendre series, the
so-called Legendre cumulants, can be determined using the
where Po(z) = 1, P1(z) = z, and m ≥1. The set of Legendre orthogonality property in Equation 4 as:
polynomials {Pm(z)}m forms a complete orthogonal basis set on
the interval [-1,1]. The orthogonality property is defined as: Lm = ∫ Pm ( z) f ( z)dz = gm ( ci ),i = 0,..., m and m = 0,1, 2,.. (7)
D
m≠n
⎧⎪ 0 , 2 where ci is the ith -order cumulant of f. Theoretically, the series 1,
∫DPm( z )Pn ( z )dx = ⎨ ,
⎪⎩ 2 m + 1 m = n
(4) with coefficients Lm calculated from Equation 7, converges to
f(z) at every continuity point of f(z) as demonstrated by Lebedev
FIG 3 - The four experimental third-order cumulant templates used in the following examples. 1- L-shape, 2 - 45°, 3-XX and 4-YY. h1 is
the distance between x and x1 and h2 is the distance between x and x2.
(1965). Finally, if only cumulants of order smaller than or equal 3. Define the template shape T for each unsampled location x0
to ω are given, then the function f(z) in Equation 1 can be using its neighbours. The conditioning data available within
approximated as follows: TEMP are then searched (Figure 5). The high-order spatial
cumulants are read from the global tree in step 1,and are
ω
~
f ( z) ≈ fω ( z) = ∑ Lm Pm ( z) (8) used to calculate the coefficients of the Legendre series.
m=
These coefficient are used to build the cpdf of Z0.
4. Draw a uniform random value in [0,1] to read from the
The above is detailed for three-dimensional spaces in conditional distribution a simulated value, Z(x0), at x0.
Mustapha and Dimitrakopoulos (2010a).
5. Add x0 to the set of sample hard data and the previously
simulated values.
6. Repeat steps 4 and 6 for the next points in the random path
defined in step 3.
7. Repeat steps 3 to 7 to generate different realisations using
different random paths.
The random path defined in step 5 concerns only the
unsampled locations. Thus, the final realisation obtained in after
step 8 honours the conditioning data.
SPATIAL CUMULANTS – EXAMPLES,

` CALCULATION AND GEOLOGICAL
INTERPRETATIONS
FIG 4 - (1) Training image; hard data locations in (2).
Examples of third-order cumulants calculated on two- and
three-dimensional data sets are presented in this section. Results
A high-order simulation method are interpreted so as to understand their use as a pattern
This section describes the high-order conditional simulation recognition tool. The data sets utilised represent complete images
method (hosim) based on spatial cumulants. A sequential and the regular grid approach described in a previous section is
procedure simulating values at unsampled locations that are followed. It should be noted that, the covariance is a measure of
randomly visited is used here. The Legendre series the periodicity between pairs of points separated by given
approximation is used to estimate the cpdfs (Mustapha and distances; similarly, the higher-order cumulant is also a measure
Dimitrakopoulos, 2010a). This expression uses Legendre of periodicity but in the direction of the symmetry axis of
polynomials which are orthogonal on the finite interval [-1,1]. template used, that is, the multiple point symmetry. In the
Then, the training images (Figure 4 (1)) and the data values examples that following, cumulants are computed on zero mean
(Figure 4 (2)) are first scaled to [-1,1]d, where d is the dimension data sets.
of the problem (ie, d = 1,2 or 3).
The hosim method first combines the training image (TI) used Spatial templates
and the samples (Figure 5) to infer the high-order spatial
In the following example, the covariance map and four
cumulants. A global calculation procedure is performed based on
directional experimental cumulants are presented, unless
a given maximal template size (TEMP) (Figure 5 (2)). This step
otherwise specified. Fourth- and fifth-order cumulants are,
consists of calculating all the spatial cumulants needed by the
respectively, computed with the and xy(-x)(-y) templates as
Legendre series approximation.
shown in Figure 6.
The main steps of hosim method are as follows:
1. Scan the training image and the sample data (Figure 5 (1)) Training image
and store the spatial cumulants in a global tree.
In Figure 7 (2) and Figure 7 (3) we consider an interpretation of
2. Define a random path visiting once all unsampled nodes. a diamond baring kimberlite pipe of the Ekati mine, NT Canada
(Nowicki et al, 2004), and its translation to a 3D binary training
(1)- Training image with hard data (2)- Template TEMP for image (76 055 nodes are used). The geological interpretation of
global calculation the pipe geometry is seen as two parts: one on the top and
another one on the bottom as shown in Figure 7. The part on the
bottom is close to a three-dimensional cone geometric shape.
The base of this cone or the intersection between the two
parts of the pipe is a horizontal section of size 150 m along
x and 200 m along y (Figure 7 and Figure 11). (For more
details, see Dimitrakopoulos, Mustapha and Gloaguen, 2010 and
Mustapha and Dimitrakopoulos, 2010b). Results for high-order
cumulants are presented first. Then, a discussion follows.
The third-order cumulant maps provide approximations of the
pipe shape as shown in Figure 8 (1) to Figure 8 (3). The fourth-
order xyz cumulants average the objects and translate them to the
origin.
For example, the pipe, considered as the only object in
Figure 7 (3), is translated to the origin as shown in Figure 9 (1).
FIG 5 - Training image in (1). The template in (2) is used for a From this figure, 2D cross-sections are shown in Figure 9 (2) and
global calculation of the spatial cumulants. Figure 9 (3). These cross-sections provide, approximately, the
1- xy(-y) 2- xy(-x)(-y)
FIG 6 - Examples of fourth-order cumulant templates (1) and fifth-order cumulant templates (2).
(2) Geological (2) Translation of the pipe

(1) Drill holes
interpretation of (1) (2) to a 3D training image
FIG 7 - A geological interpretation of a kimberlite pipe, (1) drill holes; (2) a geological interpretation of (2); (3) regular block approximation of
the pipe surrounding rock.
results obtained by the third-order cumulant maps in Figure 8. fronts from 2D cross-sections of Figure 10 are selected as shown
This conclusion is justified by the fact that the pipe shape is in Figure 12 (1.c) to Figure 12 (3.c). The third-order cumulant
approximately reflected, in a specific plane, by using orthogonal map in Figure 12 (1.a) shows a regular shape of the horizontal
shapes (xy or L-shape, xz or yz) cumulants, while it is fully sections of the pipe while the fourth- and fifth-order cumulant
characterised using xyz cumulants. These observations show, not maps, in Figure 12 (1.b) to Figure 12 (1.c), reflect some
surprisingly, the ability of the higher order cumulants to include horizontal irregularity of the pipe. The main raison comes from
key relations seen in lower orders. More generally, relations the variation of the pipe size along the vertical axis.
between cumulants can be extended for an order higher than four This variation is more described by the fourth- and the fifth-
and, in particularly, order five. The fifth-order cumulant maps are order cumulants because they manipulate points in more than two
based on four directions, and placed in four-dimensional space. directions. For example, four points are considered for the fourth
Then, cross-sections are used as detailed before. For example, order-cumulant and one of the points varies along the z direction.
Figure 10 shows a 3D cross-section of the xyz(-z) fifth-order The variations along x and y axis are less than the variation along
cumulant map. This figure translates the pipe to the origin and z axis. The third-order cumulant maps in Figure 12 (2.a) to
reflects the results of the fourth-order and, consequently, the Figure 12 (3.a) show, approximately, results close to those
third-order results are reflected too. The pipe shape varies obtained from fourth- and fifth-order maps as shown in Figure 12
strongly between the bottom and the top along the vertical axis (2) and Figure 12 (3). The fourth- and fifth-order maps detect
(z) while the variations are less along the horizontal axis (x and y) with more precision the size along x (150 m) and y (200 m) of
as shown in Figure 11. Figure 12 shows results of the third-, the intersection between the two parts of the pipe.
fourth- and fifth-order with a different way. Figure 12-(1.a) to
Figure 12-(3.a) present third-order cumulant maps using lines
contours from Figure 8. Drill hole data
Several 2D cross-sections, along x, y and z directions, are The cumulant maps calculated on the 3D training image provide
selected from the fourth-order cumulant maps in Figure 9. The good interpretation for the pipe’s shape. In the following,
(1) Third-order: xy- (3) Third-order: yz-

(2) Third-order: xz-templates
templates templates
FIG 8 - Third-order cumulants. (1) to (3) are third-order cumulants for the Fox kimberlite pipe, NT in Figure 7 (3), using xy, xz and yz
templates.
(2) Fourth-order : Three 2D cross-sections

(1) Fourth-order: xyz- templates
from (1)
FIG 9 - Fourth-order cumulants. (1) is a fourth-order cumulant map for the Fox kimberlite pipe, NT in Figure 7 (3), using xyz templates; (2)
are three 2D cross-sections from (1).
cumulants are calculated on the original data obtained from the detected then the bottom part and that because most of the data
pipe drill holes in Figure 7 (1). Figure 13 to Figure 15 show the points are in the top part, until 300 m depth, as the pipe drill holes
obtained results. In these figures, the red lines represent the show in Figure 7 (1). In Figure 13, the fourth- and fifth-order
borders of the set of high values in the cumulants maps. These cumulants maps calculated on data provide better description of
borders reflect the shape of the pipe along x and y directions the horizontal sections than the third-order map. They show
which is approximately similar to the results obtained on the 3D irregularity between 400< × <500 and along y as detected with the
training image (Figure 12). The top part of the pipe is better fourth- and fifth-order maps from the 3D training image.
(1) Fifth-order: A 3D cross-section from the (2) Fifth-order: Three 2D cross-sections from
xyz(-z) templates cumulant map at ‘-z=0’ (1)
FIG 10 - Fifth-order cumulants. (1) is a 3D cross-section at –z=0 of the five-dimensional xyz(-z) template cumulant map; (2) are three 2D
cross-sections from (1).
(1.a) xy cross-section at z=350 (1.b) xy cross-section at z=250
(2.a) xz cross-section at y=300 (2.b) xz cross-section at y=200
(3.a) yz cross-section at x=300 (3.b) yz cross-section at z=200
FIG 11 - Different horizontal and vertical 2D cross-sections of the pipe in Figure 7.
THIRD-ORDER FOURTH-ORDER FIFTH-ORDER

(1.b) 2D cross-sections (1.c) 2D cross-sections along z
r r
(1.a) {x , y} cumulant map along z from the map in from the map in Figure 10 -(1)
Figure 9 -(1)
(2.b) 2D cross-sections (2.c) 2D cross-sections along y

r r
(2.a) {x, z} cumulant map along y from the map in from the map in Figure 10 -(1)
Figure 9 -(1)
(3.b) 2D cross-sections 2D cross-sections along x from

r r
(3.a) { y, z} cumulant map along x from the map in the map in Figure 10 -(1)
Figure 9-(1)
FIG 12 - Third-, fourth- and fifth-order cumulant maps for the pipe (3D training image) in Figure 7.
xy-view
(1) Third-order (2) Fourth-order (3) Fifth-order
FIG 13 - Xy views of the third-, fourth- and fifth-order cumulant maps calculated using the data from drill holes in Figure 7 (1).
xz-view
FIG 14 - Xz views of the third-, fourth- and fifth-order cumulant maps calculated using the data from drill holes in Figure 7 (1).
CONDITIONAL SIMULATION USING CUMULANTS shown in Figures 18 (2) and (3). The developed method is also
validated by comparing the high-order statistics of the data set,
AND TRAINING IMAGES exhaustive image and the different realisations obtained. For
In this section, the simulation of a two-dimensional exhaustive example, the third-order spatial cumulant maps of the exhaustive
image (Figure 16 (1)) is shown so as to illustrate the high-order image, data set, realisations 1 and 2 are very close as shown in
conditional simulation using spatial cumulants. The example Figure 19. This last result is obtained because the new
presented herein uses 25 sample data (Figure 16 (2)), the training conditional simulation algorithm uses different cumulants orders
image in Figure 16 (3) and the conditional simulation algorithm in the Legendre series and this will guaranty the reproduction of
discussed above. not only the histogram and variograms of the sample data, but
also their high-order statistics.
Different realisations are presented as shown in Figure 17.
This figure shows that the main characteristics of the exhaustive
image are reproduced using a sparse data set (about 0.85 per cent CONCLUSIONS
of the total number of points). The 2D sections presented here
have particular and complex distributions as shown by the This paper presented developments towards a new alternative
bimodal histogram in Figure 18 (1). This figure shows the approach to modelling complex, non-linear, non-Gaussian earth
comparison between the generated realisations histograms and sciences and engineering data, as required in most applications.
the data set histogram. In addition, the realisations reproduced The new alternative framework is founded upon concepts from
the variograms along the EW and NS directions of the data set as high-order statistics that are introduced herein in a spatial context.
yz-view
FIG 15 - Yz views of the third-, fourth- and fifth-order cumulant maps calculated using the data from drill holes in Figure 7 (1).
(1)- True image (2)- 25 Sample data (3)- Training image
FIG 16 - Simulation of a horizontal 2D section of a fluvial reservoir. (1) Exhaustive image: true image, (2) 25 sample data, and (3) a
training image.
(1)- Realisation 1 (2)- Realisation 2
FIG 17 - Realisations (1) and (2) obtained by the hosim.
(1)- Histograms (2)- Variograms NS (3)- Variograms EW
γ γ
FIG 18 - Histograms (1), NS (2) and EW (3) variograms of 10 hosim realisations. The cercles refer to the data set 1 and the solid lines refer
to the realisations.
(1)- realisation 1 (2)- realisation 2
0.01 0.01
0.015 0.015
0.02 0.02
FIG 19 - Third-order spatial cumulant maps of the realisations 1 and 2, respectively.
Mathematical definitions of non-Gaussian spatial random REFERENCES

functions and their high-order spatial statistics are described in
detail, stressing the notion of spatial cumulants. The calculation of Billinger, D R and Rosenblatt, M, 1966. Asymptotic theory of kth-order
spectra, Spectral Analysis of Time Series (ed: B Harris), pp 189-232
spatial cumulants, including anisotropic experimental cumulant (John Wiley: New York).
calculations using spatial templates are introduced, and examples
Bilodeau, M and Brenner, D, 1999. Theory of multivariate statistics
of three-dimensional images presented and their characteristics (Springer: Berlin).
analysed to assess the relations between cumulants and geological Chilès, J P and Delfiner, P, 1999. Geostatistics: Modeling Spatial
patterns. The simulation of a complex image of channels is Uncertainty (John Wiley & Sons, Inc: New York).
presented using a new high-order sequential simulation method Cox, D R, 1972. Regression models and life-tables (with discussion),
which is based on the concept of high-order spatial. The results Journal of the Royal Statistical Society, Series B (Methodological),
showed a good reproduction of the main features of the exhaustive 34(2)187-220.
image using a small data set. The realisations generated Cressie, N A, 1993. Statistics for Spatial Data (John Wiley: New York).
reproduced the histogram, variogram and high-order statistics of Daly, C, 2004. Higher order models using entropy, Markov random fields
the data set. A key aspect of the simulation method based on and sequential simulation, in Geostatistics Banff, pp 215-225
spatial cumulants is the compliance of the simulated realisations (Springer).
with all statistics (any order) of the available data and avoid David, M, 1977. Geostatistical Ore Reserve Estimation (Elesevier:
possible conflicts between training images and dense data sets Amsterdam).
commonly available in mining studies. David, M, 1988. Handbook of Applied Advanced Geostatistical Ore
Reserve Estimation (Elsevier: Amsterdam).
Dimitrakopoulos, R, Mustapha, H and Gloaguen, E, 2010. High-order
ACKNOWLEDGEMENTS statistics of spatial random fields: Exploring spatial cumulants for
modelling complex, non-Gaussian and non-linear phenomena,
The work in this paper was funded from NSERC CDR Grant Mathematical Geosciences, 42(1):65-97.
335696 and BHP Billiton, as well NSERC Discovery Grant Goovaerts, P, 1998. Geostatistics for Natural Resources Evaluation
239019. Thanks are in order to Brian Baird, Peter Stone, and (Oxford: New York).
Gavin Yates, as well as BHP Billiton Diamonds and, in Guardiano, J and Srivastava, R M, 1993. Multivariate geostatistics:
particular, Darren Dyck, for their support, collaboration, data Beyond bivariate moments, in Geosatistics Tróia ’92 (ed: A Soares)
from Ekati Mine, and technical comments. volume 1, pp 133-144 (Kluver: Dordrecht).
Hohn, M E, 1999. Geostatistics and Petroleum Geology (Kluwer: New Mustapha, H and Dimitrakopoulos, R, 2010b. A new approach for
York). geological pattern recognition using high-order spatial cumulants,
Journel, A G, 1989. Five lessons in geostatistics, AGU, San Francisco. Computers and Geosciences, 36(3), pp 313-334.
Journel, A G, 1994. Modeling uncertainty: Some conceptual thoughts, Nikias, C L and Petropulu, A P, 1993. Higher–Order Spectra Analysis: A
Geostatistics for the Next Century (ed: R Dimitrakopoulos) (Kluwer: Nonlinear Signal Processing Framework, 538 p (PTR Prentice Hall:
Dordrecht). Upper Saddle River).
Journel, A G, 1997. Deterministic geostatistics: A new visit, in Nowicki, T, Crawford, B, Dyck, D, Carlson, J, McElroy, R, Oshust, P and
Geostatistics Woolongong ‘96 (eds: E Baafy and N Shofield) Helmstaedt, H, 2004. The geology of the kimberlite pipes of the
pp 213-224 (Kluwer: Dordrecht). Ekati property, Northwest Territories, Canada, Lithos. 76:1-27.
Journel, A G, 2000. Stochastic imaging of channels from seismic data Ripley, B D, 1987. Stochastic Simulation (John Wiley & Sons Inc: New
using mp statistics, SCRF, Report 13, Stanford University. York).
Journel, A G and Huijbregts, C J, 1978. Mining Geostatistics, 600 p Rosenblatt, M, 1985. Stationary Sequences and Random Fields
(Academic Press Inc: New York). (Birkhaüser: Boston), 258 p.
Kitanidis, P K, 1997. Introduction to Geostatistics – Applications in Sesay, S A and Rao, T, 1991. Difference equations for higher-order
Hydrogeology, (Cambridge University Press: New York). moments and cumulants for bilinear time series models, Jour Time
Lebedev, N N, 1965. Special Functions and Their Applications, 308 p Ser Anals, 12:159-177.
(Prentice-Hall Inc: New York). Strebelle, S, 2002. Conditional simulation of complex geological
Matérn, B, 1960. Spatial variation – Stochastic models and their structures using multiple point statistics, Mathematical Geology,
application to some problems in forest surveys and other sampling 34:1-22.
investigations, Meddelanden fran Statens Skogsforskningsinstitud, Swami, A, Giannakis, G B and Mendel, J M, 1990. Linear modeling of
49(5), Almaenna, Stockolm. multidimensional non-Gaussian processes using cumulants,
Matheron, G, 1971. The theory of regionalized variables and its Multidimensional Systems and Signal Processing, 1:11-37.
applications, Cahier du Centre de Morphologie Mathematique, Tjelmeland, H, 1998. Markov random fields with higher order
number 5, Fointanebleau. interactions, Sacndinavian Journal of Statistics, 25:415-433.
Mendel, J M, 1991. Use of high-order statistics (spectra) in signal Tjelmeland, H and Eidsvik, J, 2004. Directional Metropolis: Hastings
processing and systems theory: Theoretical results and some updates for posteriors with nonlinear likelihoods, in Geostatistics
applications, IEEE Proceedings, 79:279-305. Banff, pp 95-104 (Springer).
Mustapha, H and Dimitrakopoulos, R, 2010a. High-order stochastic Zhang, T, Switzer, P and Journel, A G, 2006. Filter-based classification
simulations for complex non-Gaussian and non-linear geological of training image patterns for spatial simulation, Mathematical
patterns, Mathematical Geosciences, 42(5)457-485. Geosciences, 38:63-80
HOME
Stochastic Mine Planning – Example and Value from Integrating

Long- and Short-Term Mine Planning Through Simulated Grade
Control, Sunrise Dam, Western Australia
A Jewbali1 and R Dimitrakopoulos2
ABSTRACT 2. minimise the potential deviations from production targets

A new multistage stochastic mine production scheduling approach is set given the available drilling and orebody models.
developed and tested in a large operating gold mine. The proposed Recent work further considers integrating market uncertainty
approach takes short-scale orebody information in the form of grade in choosing from different mine designs which also consider
control data into account. As simulated orebodies used in stochastic geologic uncertainty and show further monetary benefits of
long-term mine planning are based on sparse exploration data and while
grade control data are unavailable at the time of production scheduling,
quantifying and integrating uncertainty (Dimitrakopoulos and
the short-scale information is first simulated stochastically and then Abdel Sabour, 2007). Meagher, Abdel Sabour and Dimitra-
serves as input to the optimisation process. Stage 1 of the approach kopoulos (2010, in this volume) explore the effects of optimising
generates high density future grade control data for incorporation into the pits in a new approach that assesses the combined effect of metal
production scheduling process based on sequential co-simulation and and price uncertainties using a network flow approach; their study
pseudo cross-variograms between exploration data and grade control in shows once again the value of stochastic approaches and the
previously mined out parts of a deposit. In Stage 2, the technique of integration of key uncertainties to mine planning.
conditional simulation by successive residuals enables pre-existing The work above is all based on two elements, in addition to new
simulated orebody models to be updated using the simulated future grade stochastic mine design and scheduling methods. Firstly, the ability
control information. Stage is based on a stochastic programming mine to simulate sets of equally possible representations of the orebody
scheduling formulation that handles multiple simulated orebody models
being assessed. This set of orebodies is the input to the above
from Stage 2 and accommodates both maximising Net Present Value
(NPV) and minimising deviations from production targets. Stage 4
mentioned stochastic optimisers and represents the geological
includes quantification of risk in the produced schedules generated, uncertainty and local grade variability of the orebody, as
comparison of schedules and reporting. understood from the available drill hole data – this drastically
The application at a large operating gold mine demonstrates that the
differs from the conventional single estimated (smoothed out or
proposed approach is practical and adds value to the operation. The average type) input to conventional mine design and optimisation
approach is shown to deliver additional ore (3.6 Mt more) and metal methods and is the major reason for improved results from the
(2.6 million grams) which matches the mined reconciliations and results stochastic scheduling methods discussed. Secondly, the ability to
in a cumulative NPV which is on average A$7.7 M higher than that of a assess cash flows using evaluation methods based on multiple
stochastic schedule without the simulated grade control data and simulated realisations of the related commodity price, rather than
substantially higher (about 30 per cent) compared to the NPV from the single average and typically constant, over the LOM and metal
actual schedule of the mine. price forecasts. Both this second element and the use of
uncertainty in metal content leads to improvements in the last two
INTRODUCTION studies above.
A limit of both stochastic and conventional long-term planning
Stochastic mine planning is a relatively recent development aimed and optimisation of production planning is that short-term
at addressing uncertainty in ore supply from an orebody, schedules may frequently deviate from the expectations of
commodity prices and metal demand, as well as expanding to long-term ones. While long-term production scheduling is used to
other issues of uncertainty in mine planning. The existing work maximise the Net Present Value (NPV) of a project, short-term
focuses largely on open pit mine design and production scheduling production scheduling focuses on meeting short-term production
with uncertain grades/metal content and geological conditions. It demand within the long-term plan, given the processing capacity
includes life-of-mine (LOM) production scheduling based on of the mill and managing the quality of ore being processed. Both
optimal mining rates and simulated annealing (Godoy and short- and long-term scheduling are usually based on
Dimitrakopoulos, 2004), simulated annealing with constant exploration-scale orebody models, whereas the actual short-term
mining rates (Leite and Dimitrakopoulos, 2007), stochastic integer production performance is controlled by local, blasthole-scale data
programming formulations including a stochastic stock pile used for grade control and ore/waste selection, usually not
(Ramazan and Dimitrakopoulos, 2009, 2007), further tests of this available at the time of scheduling. As a result, short-term
last approach (Leite and Dimitrakopoulos, 2010, in this volume), production sequences deviate from the long-term plans and are
expansion for multiple elements (Benndorf and Dimitrakopoulos, adjusted to meet mill demand or the production performance may
2010, in this volume) and others. The two key aspects of all the substantially deviate from forecasts. In short, discrepancies occur
above mentioned approaches and related example case studies are between what is forecasted and what is actually mined.
that the derived long-term production schedules: A solution to the above is to consider production scheduling
1. have substantially higher present value than the tradition- approaches that are capable of reflecting short-scale behavior of
ally used approaches with differences between 20 per cent the orebody beforehand, that is, at the planning stage and through
and 30 per cent, and grade information available before the actual grade control
drilling. High density grade control information offers tighter
controls on geology and mineralisation characteristics and predicts
1. GAusIMM, Resource Estimation Geologist, Rio Tinto, 152-158 qualities and quantities closer to what is actually being mined.
St Georges Terrace, Perth WA 6000. Email: [email protected] Although not available prior to blasthole drilling, grade control
2. FAusIMM, Professor and Director, COSMO – Stochastic Mine data can be simulated in different ways. Simple random sampling
Planning Laboratory, Department of Mining and Materials errors added to simulated realisations from exploration data, have
Engineering, McGill University, Montreal QC H3A 2A7, Canada. been used to assess production schedules reported in a feasibility
Email: [email protected] study (Guardiano, Parker and Isaaks, 1995) and a conceptually
A JEWBALI and R DIMITRAKOPOULOS
similar more sophisticated approach is considered for assessing As a result of the ability to link information at different scales
mineral resources by Journel and Kyriakidis (2004). In operating in space and time, short scale variability, short- and long-term
mines, errors from field duplicates or from nearest paired grade schedules are integrated. The proposed approach or process has
control holes and drill holes have been used to simulate four stages shown in Figure 1, which also shows how this
realisations of future grade control data and assess short-term compares to the four steps followed in the existing stochastic
production schedules at Escondida copper mine, Chile LOM planning:
(Khosrowshahi, Shaw and Yeats, 2007). Multiple simulations of
correlated sampling errors using data from previously mined parts 1. Stage 1: Generate high density future grade control data for
of the deposit have been used at Morila gold deposit, Mali (Peattie incorporation into the production scheduling process. This
and Dimitrakopoulos, 2009) with excellent results in data is not yet available at the time of planning and is
reconciliations. With the exception of the last one, past approaches simulated.
are relatively simplistic. All existing work in this area assesses 2. Stage 2: Update pre-existing simulated orebody models
potential risks in production schedules or potential reserves – using the simulated future grade control information and a
however, it does not address the major question: How to improve suitable stochastic simulation technique.
or reduce deviations of forecasts from actual production at the
time of planning while maximising the present value of an asset. 3. Stage 3: Generate production schedules using a stochastic
This can be effectively addressed only by technically integrating optimisation method handling multiple simulated orebody
short-scale orebody variability and grade control with, short- and models from Stage 2, which accommodates both maximising
long- term LOM production planning into one approach. net present value and minimising deviations from production
Short-scale information is the core element in the targets.
above-mentioned integration. As this information is unavailable 4. Stage 4: Quantify grade risk in the produced schedules,
and needs to be stochastically simulated at the time of LOM compare schedules with and without simulated short scale
planning, the way to utilise a range of possible scenarios is orebody information, and report mine schedules as needed.
through the stochastic optimisation formulations used in
These four stages are discussed next in greater detail.
stochastic mine planning as reviewed above. This paper
contributes a new multistage approach to production scheduling
that incorporates, short-scale deposit information and related Existing Stochastic LOM Proposed Multistage Approach with
grade uncertainty into the scheduling process. The approach Scheduling Process Short-scale Information
allows for the realistic integration of short- and long-term mine
production schedules, as well as the generation of more reliable
mine production forecasts. A case study at an operating gold Stage 1
deposit demonstrates the approach, provides comparisons Simulation of orebody models from Simulation of high density future
exploration data grade control data
between the traditional schedules of the mine, accounts for
stochastic schedules before and after simulated dense grade
control information and quantifies the expected monetary value Stage 2
of the method. In the next sections, the proposed approach is first Updating of the existing orebody
models with the future data
outlined, the case study presented, results are analysed and
conclusions follow.
Stage 3
A NEW MULTISTAGE APPROACH TO Stochastic optimisation and Stochastic optimisation and generation
PRODUCTION SCHEDULING generation of production schedules of production schedules
The multistage approach to production scheduling proposed here

is as sequence of steps employing separate techniques for each Stage 4
stage. The approach differs conceptually from the traditional Schedules
approaches in many aspects. A major difference is that it requires Quantification of risk and analysis in schedule
multiple, equally probable representations of the orebody, which
are generated from spatial Monte Carlo simulation methods and
at two different levels. The first level is that of exploration-type
data sets and information; the second level is that of grade
control data. This simulation-based framework assists in
quantifying and generating risk managing schedules as well as
accommodates the interaction of information at different scales,
both in space and time. Different scales in space refer to the:
FIG 1 - Schematic presentation of the proposed multistage mine
• local variability of the deposit and local classification of scheduling approach.
materials selected as ore and waste through a grade control
process, and
Stage 1 – simulation of future grade control
• physical parts of the orebody that are mined over long
periods of time as individual pit units (cut-backs) and parts of
drilling data
individual pit units that correspond to the parts of cut-backs High density future grade control data can be simulated using
mined over short periods of time from the same mining front. several approaches, such as those reviewed in the introduction. A
Different scales in time refer to the: more elaborate and effective method is outlined here and is used
in the case study that follows. The method, schematically shown
• yearly time units on which the long-term plans are reported, in Figure 2, assumes that there are mined out parts of an orebody
and with grade control information of comparable quality and
• relatively shorter time periods in short-term planning characteristics to the remainder (as yet un-mined) of the orebody
(weekly, monthly or quarterly). being studied.
STOCHASTIC MINE PLANNING – EXAMPLE AND VALUE FROM INTEGRATING LONG- AND SHORT-TERM MINE PLANNING
where:
Bench/section of pit already mined out
P is the number of periods to schedule
Exploration data
Grade control data N is the total number of blocks to schedule
S refers to the number of simulated orebody models
R is the number of targets
Define relationship
t
X i is a binary variable indicating block i mined in period
t (1 = mined/0 = not mined)
E(NPV)ti is the expected NPV to be generated if block i is
Exploration data mined in period t
Simulate grade control data
Yu tr is the excess amount produced compared to the target
Bench/Section of pit NOT yet mined out (grade/tonnage) r in period t for orebody model s
s
Cu tr is the unit cost to penalise Yu tr in period t
FIG 2 - Simulating future grade control data. YItr is the deficient amount produced in relation to target
(grade/tonnage) r in period t for orebody model s
Exploration drilling and grade control data from a mined out s
CItr is the unit cost to penalise YItr in period t
section or bench of a deposit is used to quantify the spatial
The first component in Equation 1 maximises the total
relationship between the two types of data. As it is not physically
discounted cash flow and the expected Net Present Value (NPV)
possible to have exploration and grade control information
is determined by calculating economic values for each simulated
available at the same location, their spatial cross-correlation
model, then averaging them. This component reflects the
(relationship) is quantified through the so-called pseudo
conventional goal of optimising total discounted cash flow over
cross-variogram (Myers, 1991). Having quantified this spatial
the LOM. The second component minimises the deviations
relationship, a method such as the sequential Gaussian
between expected productions as these are described by
co-simulation of two variables (Goovaerts, 1997) can be used to
simulated orebody models used as input and a set of mine
generate the future grade control data in un-mined sections of an
production targets and minimises the risk of not meeting targets
open pit. The co-simulation approach ensures that the simulated
due to grade uncertainty. The deviations are calculated over the
future grade control data have the same distribution as past grade
set of simulated orebody models and are penalised with costs.
control information and the same auto- and cross-correlation
This second component reflects the goal of short-term
(with exploration data).
production scheduling, and its use together with the first
component provides the means to integrate long- and short- term
Stage 2 – updating of existing simulations with planning, in the context of the approach proposed herein. It
future grade control data should be noted that a novelty of this formulation is that the costs
The second stage of the production scheduling approach presented can be discounted for each subsequent period using a geological
herein involves the updating of pre-existing simulated orebody discount rate. As a result, the cost related to not meeting mill
models. This is because orebodies under study have frequently requirements can be set higher in the first period than in
already been simulated for other purposes, or because there can be subsequent periods, allowing the risk distribution between
more than one scenario of possible sets of future data to consider, production periods to be managed.
and so on. Generally the ability to update pre- existing realisations
of an orebody is useful in improving the efficiency of the process. Stage 4 – quantification of uncertainty
The only known simulation technique that is capable of providing
updating capabilities to accommodate the updating of pre-existing Stage 4 aims to quantify grade risk in the produced schedules
simulations with new data such as grade control is the technique of that have been generated. Any schedule can be assessed against
conditional simulation by successive residuals or CSSR possible scenarios of orebodies, be they constructed from
(Vargas-Guzman and Dimitrakopoulos, 2002). The end result of exploration datasets only or in combination with future grade
the use of CSSR is a set of simulated orebody models that are control data. This allows comparisons and assessments, including
conditioned on both the known exploration data and the simulated an evaluation of effects from potential short-scale orebody
future grade control data. Note that new information for updating variability.
may also be additional data from near mine exploration drilling, or
in fill drilling. CASE STUDY AT A GOLD MINE
Stage 3 – stochastic integer programming The gold deposit and mine
formulation
The gold deposit in this case study lies along a shear zone and
To generate production schedules, a stochastic optimisation the general trend of the gold mineralisation is parallel to this
formulation based on Stochastic Integer Programming (SIP) is zone. Gold mineralisation can be both secondary and primary.
used, and it is a simpler version of the one developed in Ramazan Secondary mineralisation is related to iron oxidation fronts and
and Dimitrakopoulos (2007, 2010) as well as discussed in Leite water tables and to aggregation within paleochannels, while
and Dimitrakopoulos (2010, in this volume). The objective primary mineralisation can roughly be related to high strain shear
function of this formulation maximises the expected discounted zones by low strain stockwork-vein zones. Based on the deposit
cash flows while minimising the cost of deviating from the set geology, 17 domains are inside the open pit (sections of which
production targets and it is: are shown in Figure 3). Mining is done using four excavators
which load into 25 rear dump trucks. Since opening, the mine
P N S P R has produced more than two million ounces of gold. The pit is
Max ∑ ∑ E(NPV)ti Xti − ∑ ∑ ∑ ( s Cu trYu tr + s Cl trYl tr ) (1) about 2.1 km long and 1.1 km wide with a final depth of 450 m
t = 1 i= 1 s = 1 t = 1 r= 1 to be reached at end of its life. The total amount of waste mined
Generation of future grade control data

2200 Within the mined out section of the pit, 75 789 grade control drill
North
P1 holes were available. These were separated based on the geological
domain they belonged to. In addition, 3934 exploration diamond
and reverse circulation holes were available within the same section
of the pit. These were also divided based on geological domains.
P2 For each geological domain these two sets of information were used
P3 to derive the relationship between exploration and grade control
data. This relationship can than be extended to unmined sections of
P4 the pit where only exploration data is available. For the purposes of
2000 this study, twenty future grade control data simulations were
Y=100089 generated, using the sequential cosimulation method discussed
earlier, on 5 m × 7 m × 2.5 m spacing in the unmined sections of
the pit (the volume displayed in Figure 3).
2200 P1
Updating existing simulations with future grade
control data
Stage 2 of the production scheduling approach involves using
P2 conditional simulation by successive residuals to update pre-
P3 existing orebody models. The 20 available simulations (on 2 m ×
2 m × 2.5 m spacing) conditioned on exploration data were
updated with the 20 simulated future grade control data generating
P4 a set of 400 simulations that are conditioned on both the known
Y=99824 exploration data and the simulated future grade control data
2000 scenarios. Twenty updated simulations are chosen at random for
P5
further study. For reasons of comparison and processing, the
simulations are reblocked to 15 × 30 × 7.5 m3 blocks, which is the
2200 block size used at the mine. Grade tonnage curves are shown in
South Figure 4. It is apparent from the figure that for the model based on
P1
exploration data, the average grade above cutoff is higher at all
P2 cutoffs. The amount of tonnes above cutoff is lower compared to
the updated models.
P3
Grade tonnage curves
P4 12
Millions
100.0
Y=99599 10
2000
80.0
8
Grade g/t
Tonnage
50200 50400 60.0

6
Period (years)
40.0
4
1 2005
20.0 2
2 2006
3 2007 0.0 0
0 0.4 0.8 1.2 1.6 2 2.4 2.8 3.2 3.6 4 4.4 4.8
4 2008 Cut-off grade g/t
5 2009
FIG 4 - Grade tonnage curves for the models based on exploration
data (dotted lines) and the models updated with simulated future
grade control data (solid lines).
FIG 3 - East-West sections of the yearly schedule in a gold mine.
Figure 5 displays the ore tonnes and metal recovered when

in 2004 was about 43 Mt, which drops to about 26 Mt in 2005 mining is done according to the mine’s yearly schedule (as
and again to 11 Mt in 2007. This will necessitate a resizing of the
shown in Figure 3) when short scale deposit information is taken
equipment fleet sometime in mid-2005. Figure 3 depicts the
long-term (yearly) production schedule in use at the mine. The into account. The risk profiles were generated by running the 20
mine’s schedule proceeds one layer at a time and indicates that updated models through the mine’s yearly schedule. During this
the pit will be exhausted by the end of the fifth year. At this process the tonnes of ore/metal produced and NPV generated by
mine, blasting occurs on 7.5 m benches and grade control is done each updated model was recorded and plotted on Figure 6. The
using reverse circulation drilling on 5 m × 7 m spacing. The plant updated models indicated that the amount of ore tonnes produced
capacity is 3.6 Mt/a and the mill circuit consists of crushing and is not enough to fill the plant to capacity (3.6 Mt/a) for all the
grinding after which the gold is recovered through a Carbon in years from 2005 to 2009. They also indicate a 50 per cent chance
Leach (CIL) circuit. of negative NPV in 2005 when insufficient metal is produced.
Risk profile for ore production for Sunrise Dam long-term schedules derived using orebody models based on two sets of
schedule; updated models information. As a result, it provides a look at the consequences of
5.0 not taking closer spaced grade control information into account
Millions
4.5 when performing production scheduling.

4.0 The production schedules were generated for four month
3.5 periods (short-term schedules). The production capacity for each
four month period was set at 1.2 MT and the targeted grade
Tonnage
3.0
2.5 constraints between two and 7 g/t. To make the schedules practical
2.0 for equipment mobility and space, the schedules were smoothed.
1.5 After smoothing, the simulated orebody models are run through
1.0 their respective schedules to generate risk profiles for ore, metal
0.5 and NPV. Each short-term schedule was regrouped to generate the
0.0 yearly LOM schedule of the mine. The risk profiles of the
2004 2005 2006 2007 2008 2009 2010 production schedule derived from the 20 exploration based models
Years for ore, metal and NPV are shown in Figure 6. As the figure
Risk profile for ore production
Risk profile of metal content for Sunrise Dam long-term schedule; Schedule derived from simulations based on exploration data
updated models 4.3
Millions grams
Millions
18
4.1
16
3.9
14 3.7
Tonnage
12 3.5
Metal content
10 3.3
8 3.1
6 2.9
4 2.7
2
2.5
2004 2005 2006 2007 2008 2009
0 Periods (years)
2004 2005 2006 2007 2008 2009 2010
Years
Millions grams
Risk profile for metal production

Schedule derived from simulations based on exploration data
Risk profile of NPV for Sunrise Dam long-term schedule; 22
updated models 20
18
250.0
Millions AUD
16
Metal content
14
200.0
12
10
150.0 8
6
NPV
100.0 4
2
50.0 0
2004 2005 2006 2007 2008 2009
Periods (years)
0.0
2004 2005 2006 2007 2008 2009 2010 Risk profile for NPV
-50.0 Schedule derived from simulations based on exploration data
Years 250.0
Millions AUD
200.0
Mill target (ore production)
Average of the simulations 150.0
NPV
Simulations 100.0
FIG 5 - Performance of the mine’s yearly schedule when local 50.0

orebody information is taken into account.
0.0
2004 2005 2006 2007 2008 2009
Deriving production schedules using the SIP Periods (years)
formulation
The following section describes the process, as given in Figure 1, Mill target (ore production)
used to derive the production schedules. First the 20 simulated Average of the simulations
orebody models conditioned on exploration data are put through
the SIP formulation in Equation 1. Next, the 20 updated orebody Simulations
models are then run through the SIP formulation and a production
schedule accounting for short-scale orebody information is FIG 6 - Risk profile of ore, metal and NPV for the production
derived. This methodology allows for a comparison of the schedules derived from orebody models based on exploration data.
shows, the yearly production target of 3.6 Mt/a is met for 2005 to exploration data (that do not account for local orebody
2007. Note that this schedule predicts that the pit will be knowledge and characteristics) underestimate the actual tonnes
exhausted by 2008. Figure 7 describes the schedule derived from of ore mined from the orebody. The derived exploration based
the updated models. This schedule extends the life of mine by schedules are a conservative option. Taking local grade control
another year to 2009. As the figure shows, the short-term schedule information into account indicates that there is probably
has no problem delivering 3.6 Mt/a of ore in all years except in sufficient ore to extend the life of mine to 2009 compared to
2009, when the last of the ore will be mined. 2008 when only exploration data is used. Figure 8 displays the
The schedules based on updated models produce 3.6 Mt more production schedules for the exploration based simulations and
ore and as a consequence 2.6 million grams more metal, which the updated simulations. As Figure 8 shows, the two schedules
results in a cumulative NPV that is A$7.7 M higher. The are different in that different sections of the pit are mined out in
production schedules based on orebody models derived from different time periods. However, when compared to Figure 3,
both of the SIP derived schedules mine the pit out in a series of
successive cones, while the mine’s yearly schedule proceeds in
Risk profile for ore production
Schedule derived from updated models layers.
4.7
Millions
4.5
4.3
TABLE 1
4.1 Cumulative averages for ore tonnes, metal and NPV over all
3.9 simulated orebody models (both exploration data based and
Tonnage
3.7 updated models).

3.5
3.3
SIP and SIP and Mines schedule
simulations updated performance
3.1
with simulations with updated
2.9
exploration with ‘future’ simulations
2.7
data data
2.5
2004 2005 2006 2007 2008 2009 2010 Ore tonnes 14 18 10
Periods (years) (million grams)
Metal tonnes 52 55 38
Risk profile for metal production (million grams)
Schedule derived from updated orebody models
18 NPV 552 560 330
Million grams
(million AUD)
16
14
12
Metal content
10 2200
P4 P1
8
P2 P1 P3
6
P5 P4
4
2000 Y=100089
North
2
0
2004 2005 2006 2007 2008 2009 2010 P5
2200
Periods (years)
P1
P1
P2
Risk profile for NPV P3
Schedule derived from updated models P2
2000 Y=99824 P3
Millions AUD
200.0
P4
P4
150.0
2200
P3 P3
P1
NPV
P5 P2
100.0
Y=99599 P4
2000 South
50.0
50200 50400 50200 50400
0.0 Period (years)

2004 2005 2006 2007 2008 2009 2010
Periods (years)
1 2005
2 2006
3 2007
Mill target (ore production)
4 2008
Average of the simulations
5 2009
Simulations
FIG 7 - Risk profiles of ore, metal and NPV for the production FIG 8 - West-east sections of the Sunrise pit displaying production
schedule derived from orebody models displaying grade control schedules derived from simulation based on exploration data
characteristics. (right) and updated models (left).
CONCLUSIONS Godoy, M and Dimitrakopoulos, R, 2004. Managing risk and waste

mining in long-term production scheduling of open pit mines:
A new multistage stochastic mine production scheduling Transactions of Society for Mining, Metallurgy and Exploration,
approach presented herein includes four stages: 316:43-50.
Goovaerts, P, 1997. Geostatistics for Natural Resources Evaluation, 483 p
1. Stage 1: generates high density future grade control data for (Oxford University Press: New York).
incorporation into the production scheduling process. It is Guardiano, E, Parker, H M and Isaaks, E H, 1995. Prediction of
based on sequential Gaussian co-simulation of exploration recoverable reserves using conditional simulation: A case study for
and grade control data from previously mined out parts of a the Fort Knox gold project, Alaska, Unpublished report, Mineral
deposit. Resources Development Inc, Foster City.
2. Stage 2: employs conditional simulation by successive Journel, A G and Kyriakidis, P C, 2004. Evaluation of Mineral Reserves:
A Simulation Approach, 216 p (Oxford University Press: New York).
residuals to update pre-existing simulated orebody models
with the simulated future grade control information from Khosrowshahi, S, Shaw, W J and Yeates, G A, 2007. Quantification of
risk using simulation of the chain of mining – A case study on
Stage 1. Escondida Copper, in Orebody Modelling and Strategic Mine
3. Stage 3: uses a stochastic integer programming mine Planning, second edition (ed: R Dimitrakopoulos), pp 33-42 (The
scheduling formulation and balances both maximising net Australasian Institute of Mining and Metallurgy: Melbourne).
present value and minimising deviations from production Leite, A and Dimitrakopoulos, R, 2007. A stochastic optimisation model
targets. for open pit mine planning: Application and risk analysis at a copper
deposit, Transactions of the Institutions of Mining and Metallurgy,
4. Stage 4: includes quantification of risk in the produced Mining Technology, 116:A109-A118.
schedules generated and reporting. Leite, A and Dimitrakopoulos, R, 2010. Production scheduling under
The application at a large operating gold mine demonstrates programming at an open pit copper mine and comparison to
that the proposed approach is practical and adds value to the conventional scheduling, in Advances in Orebody Modelling and
operation. The approach is shown to deliver 3.6 Mt of additional Strategic Mine Planning I (ed: R Dimitrakopoulos), pp 35-40 (The
ore and 2.6 million grams more metal – which matches better Australasian Institute of Mining and Metallurgy: Melbourne).
with the mine’s reconciliations. The approach also results in a Meagher, C, Abdel Sabour, S A and Dimitrakopoulos, R, 2010. Pushback
cumulative NPV which is on average $7.7 M higher that that of a design of open pit mines under geological and market uncertainties,
comparable stochastic schedule without the simulated grade in Advances in Orebody Modelling and Strategic Mine Planning I
control data and substantially higher when compared to the NPV (ed: R Dimitrakopoulos), pp 291-298 (The Australasian Institute of
from the actual schedule of the mine. Mining and Metallurgy: Melbourne).
Myers, D, 1991. Pseudo cross-variograms, positive definitness and
cokriging, Mathematical Geology, 14:805-816.
ACKNOWLEDGEMENT Peatie, R and Dimitrakopoulos, R, 2009. Forecasting recoverable ore
The authors are grateful to AngloGold Ashanti for funding and reserves and their uncertainty at Morila gold deposit, Mali – An
efficient simulation approach and future grade control drilling, in
multifaceted support, Rio Tinto for funding, as well as additional
Proceedings Orebody Modelling and Strategic Mine Planning (ed: R
support from NSERC grants and McGill’s COSMO Stochastic Dimitrakopoulos), pp 193-200 (The Australasian Institute of Mining
Mine Planning Laboratory. and Metallurgy: Melbourne).
Ramazan, S and Dimitrakopoulos, R, 2007. Stochastic optimisation of
REFERENCES long-term production scheduling for open pit mines with a new
integer programming formulation, in Orebody Modelling and
Benndorf, J and Dimitrakopoulos, R, 2010. Stochastic long-term Strategic Mine Planning, second edition (ed: R Dimitrakopoulos),
production scheduling of iron ore deposits – Integrating joint pp 385-392 (The Australasian Institute of Mining and Metallurgy:
multi-element geological uncertainty, in Advances in Orebody and Melbourne).
Strategic Mine Planning I (ed: R Dimitrakopoulos), pp 151-158 (The Ramazan, S and Dimitrakopoulos, R, in press. Production scheduling
Australasian Institute of Mining and Metallurgy: Melbourne). with uncertain supply – A new solution to the open pit mining
Dimitrakopoulos, R and Abdel Sabour, S A, 2007. Evaluating mine plans problem, European J Operations Research and COSMO Lab
under uncertainty: Can the real options make a difference? Resources research report no 1, 2007, pp 257-294.
Policy, 32:116-125. Vargas-Guzman, J A and Dimitrakopoulos, R, 2002. Conditional
Dimitrakopoulos, R, Martinez, L S and Ramazan, S, 2007. A maximum simulation of random fields by successive residuals, Mathematical
upside/minimum downside approach to the traditional optimisation Geology, 34:507-611.
HOME
Author Index
A E L
Abdel Sabour, S A 29, 291 Elkington, T 241 Leenders, B 113
Acuña, E 263 Leite, A 35
Allen, L 193, 199 Lopes, J A 205
F
Arvidson, H 43
Fava, L 263
M
Fereira, J 79
B Fernandes, J B 205 Malajczuk, W 101
Bandarian, E M 79 Filippov, P A 129 Meagher, C 291
Benndorf, J 151, 167 Flitton, T 257 Mueller, U A 79
Bodon, P 277 Freidin, A 249 Mustapha, H 309
Botelho, M C 123 Fricke, C 277 Myers, P 241
Boucher, A 59
Boyle, C 67
N
Bray, A 223
G
Neverov, A A 129
Brazil, M 235 Gloaguen, E 91
Neverov, S A 129
Brisebois, K 299 Godoy, M 21, 215
Groeneveld, B 113
O
C
Oparin, V 249
Carter, P 271 H
Osterholt, V 43
Chitiyo, G 223 Hall, S 263
Chou, C L 215 Hammitt, J 299
Coutts, B 143 Hardcastle, S 263
P
Hardtke, W 193, 199 Peattie, R 257
Hart, W 299 Perry, B 193, 199
D
Herod, O 43
de Carvalho Pereira, G J 123
Hilário, J A 123 R
Dimitrakopoulos, R 13, 29, 35
91, 151
Rendu, J-M 173
215, 291
309, 321 J Reynolds, K 159
Dincer, T 159 Jewbali, A 321 Richardson, S 79
Douglas, I 193, 199 Richmond, A 285
Dowd, P A 51 Robins, S 179
Du Pisani, P 223 K Rocchi, L 271
Dunn, P 263 Kemp, C 223 Rodriguez, P C 123
Durham, R 241 King, B 7 Rosas, C F 205
S Topal, E 113 X
Sandeman, T 277 Xu, C 51
Silva, E A 123 V
Stanford, C 277 van der Riet, J 101
Z
Stone, P 101, 271 Vanzela, G A 205
Zuckerberg, M 101
Verly, G 299
Volz, M G 235
T
Tapsiev, A 249
Thomas, D A 235 W
Tomaselli, B Y 123 Wharton, C 107
Tonkin, C 135 Whittle, J 3

Advances in Orebody Modelling and Strategic Mine Planning I PDF

Uploaded by

Copyright:

Available Formats

Advances in Orebody Modelling and Strategic Mine Planning I PDF

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Advances in Orebody Modelling and Strategic Mine Planning I PDF

Uploaded by

Copyright:

Available Formats

Advances in

Orebody Modelling and

The Australasian Institute of Mining and Metallurgy

All papers published in this volume were refereed prior to publication.

ISBN 978 1 921522 21 5

Desktop published by:

Compiled on CD ROM by:

Enrique Acuna Ute Mueller

Jorg Benndorf Hussein Mustapha

Alexandre Boucher Mike O‘Brien

Marcus Brazil Harry Parker

Peter Carter Richard Peattie

Snehamoy Chatterjee Salih Ramazan

Michel Dagbert Jean-Michel Rendu

Roussos Dimitrakopoulos Andrew Richmond

Tamer Dincer Sabry Abdel Sabour

Peter Dowd Michael Samis

Marcelo Godoy Peter Stone

Nicole Grieco Olivier Tavchandjian

John Hadjigeorgiou Erkan Topal

Geoff Hall Marcus Volz

Xu Huan Chris Wharton

Arja Jewbali David Whittle

Brett King Jeff Whittle

Amina Lamghari Gavin Yeates

This volume comprises 41 papers under the following topics:

The Global Optimiser Works – What Next? J Whittle 3

Optimal Mining Principles B King 7

Stochastic Mine Planning – Methods, Examples and Value in an R Dimitrakopoulos 13

A Risk Analysis Based Framework for Strategic Mine Planning M Godoy 21

Production Scheduling Under Metal Uncertainty – Application of A Leite and R Dimitrakopoulos 35

Modelling Orebodies and Applications

Regional Three-Dimensional Modelling of Iron Ore Exploration V Osterholt, O Herod and 43

Conditional Simulation of Grades Controlled by Geological C Xu and P A Dowd 51

Conditional Simulation Methods to Determine Optimum Drill Hole C Boyle 67

Transformation Methods for Multivariate Geostatistical Simulation E M Bandarian, U A Mueller, 79

Conditional Wavelet Based Simulation of Non-Stationary E Gloaguen and 91

Advancing Strategic Mine Planning – A Software Perspective C Tonkin 135

Large-Scale Integrated Applications

Integrated Strategic Planning at EKATI Diamond Mine B Coutts 143

Stochastic Long-Term Production Scheduling of Iron Ore J Benndorf and 151

Overall Optimisation of Yerilla Project T Dincer and K Reynolds 159

Investigating the Variability of Key Coal Quality Parameters in J Benndorf 167

A Geology-Based Joint Conditional Simulation of a Lateritic J A Lopes, C F Rosas, 205

Resource/Reserve Classification with Integrated Geometric and R Dimitrakopoulos, M Godoy 215

Optimisation is Moving Underground

The New World and Strategic Innovation

Geologically Enhanced Simulation of Complex Mineral Deposits H Mustapha and 309

Author Index 329

Copies of all books currently in print can be obtained from

The Institute office - Tel (03) 9662 3166

Key: • Out of print

Copies of all books currently in print can be obtained from

The Institute office - Tel (03) 9662 3166

Key: • Out of print

CONFERENCE, SYMPOSIUM AND MISCELLANEOUS

5. Towards 2000 – National Conference on Ironmaking Resources and Reserves 1997

The Global Optimiser Works – What Next?

HOW THE PROBER SERIES OF PROGRAMS