See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/268584875

New Developments To Rapidly Delineate 3D Iron Ore Geology And Resource

Models

Conference Paper · January 2009

DOI: 10.1190/segj092009-001.18

CITATION READS

1 188

5 authors, including:

Desmond James FitzGerald Antonio Guillen

Intrepid Geophysics, Australia French Geological Survey
123 PUBLICATIONS 655 CITATIONS 59 PUBLICATIONS 1,824 CITATIONS

SEE PROFILE SEE PROFILE

Horst Holstein Asbjorn Norlund Christensen

Aberystwyth University Nordic Geoscience
92 PUBLICATIONS 877 CITATIONS 39 PUBLICATIONS 310 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Desmond James FitzGerald on 21 February 2019.

The user has requested enhancement of the downloaded file.

Delineate Three-Dimensional Iron Ore Geology and Resource
Models Using the Potential Field Method
D FitzGerald1, J-P Chilès2 and A Guillen3
ABSTRACT
Most 3D geological modelling tools were designed for the needs of the oil
industry or detailed mine planning and are not suited to the variety of
situations encountered in other application domains. Moreover, the usual
modelling tools are not able to quantify the uncertainty of the geometric
models generated. The potential field method was designed to build 3D
geological models from data available in geology and mineral exploration,
namely the geological map and a digital terrain model (DTM), structural
data, borehole data and interpretations of the geologist. This method
considers a geological interface as a particular isosurface of a scalar field
defined in the 3D space, called a potential field. The interpolation of that
field, based on cokriging, provides surfaces that honour all the data. The
3D model and its parts are always consistent with the observations.
New developments allow the covariance of the potential field to be
identified from the structural data. This makes it possible to associate
sensible cokriging standard deviations to the potential field estimates and
to express the uncertainty of the geometric model. It also, for the first
time, gives a statistically optimal, geologically sound way of interpolating
geology, other than directly joining the dots as you do with CAD.
Practical implementation issues for producing 3D geological models
are presented: how to handle faults, how to honour borehole ends, how to
take relationships between several interfaces into account, how to model
thin beds over many kilometres, how to optimise lithological properties
and how to integrate gravimetric and magnetic data.
We describe all geology surfaces and volumes using implicit functions.
These are then rendered onto the required sections, plans, etc. The
estimation of orebody grades and tonnes, using an unbiased and optimal
geostatistical technique, makes use of the stratigraphically bound 3D
geology model.
An application to the geological modelling of the Hamersley iron ore
district, Australia, is briefly presented.
INTRODUCTION
The resource evaluation of an iron ore deposit is often performed
in three steps:
1. delimitation of the boundaries of the units corresponding to
the various geological formations or ore types;
2. estimating densities; and
3. estimation of grades within each unit. In simple cases (eg a
series of subhorizontal layers), the geometric model can be
built using 2D geostatistical techniques (kriging or
cokriging of the elevations or thicknesses of the various
horizons) which also quantify the uncertainty of the model.
A recent paper by Osterholt, Herod and Arvidson (2009) shows
these steps. A lot of effort has been undertaken to develop 3D
modelling tools capable of handling more complex situations (eg
Mallet, 2003). Most of them were designed to fulfil the needs of
the oil industry, namely for situations where the underground
model can be mostly defined from seismic data. Deterministic
methods are also available to interpolate between subparallel
interpreted cross-sections.
1. FAusIMM, Director, Intrepid Geophysics, 2/1 Male Street, Brighton
Vic 3186. Email: des@intrepid-geophysics.com
2. Ecole Des Mines de Paris, Fontainebleau 77305, France.
Email: jean-paul.chiles@ensmp.fr
3. Intrepid Geophysics, 2/1 Male Street, Brighton Vic 3186.
Email: antonio@intrepid-geophysics.com
Iron Ore Conference Perth, WA, 27 - 29 July 2009
When assessing resources, knowledge of the degree of
uncertainty of the estimation is as important as the estimate
itself. Uncertainty on the boundaries and volumes of the various
units is often a major part of the global uncertainty. When 2D
geostatistical techniques can be used, the quantification of that
uncertainty by an estimation variance is a valuable by-product of
the estimation process (Chilès and Delfiner, 1999). In contrast
usual 3D modelling tools are not able to quantify the uncertainty
attached to the interpolated model, whereas that uncertainty can
be quite large.
The potential field method (Calcagno et al, 2008) was
designed to build 3D geological models from data available in
geology and mining exploration, namely:
• a geological map and a digital terrain model (DTM),
• structural data related to the geological interfaces,
• borehole data,
• gravity data, and
• interpretations from the geologist.
It is not limited to sedimentary deposits and does not require
seismic data (such data would be useful but are seldom available
in geological, mining, and civil engineering applications).
The potential field method defines a geological interface as an
implicit surface, namely a particular isosurface of a scalar field
defined in the 3D space – the potential field. The 3D
interpolation of that potential field, based on cokriging, provides
isosurfaces that honour all the data. Recent developments allow
the covariance to be determined from the structural data, which
makes it possible to associate sensible cokriging standard
deviations to potential field estimates and to translate them into
uncertainties on the 3D model.
In Appendix A, we cover the basic principle of the method,
present the inference of the potential field covariance from the
structural data, and explain how the uncertainty of the 3D model
can be quantified. In the body of this paper we examine several
practical issues: how to form a covariance matrix, how to handle
faults, how to incorporate lithology property distributions, how to
take relationships between several interfaces into account, how to
link 3D geometrical modelling and inverse modelling of gravi-
metric and magnetic data. We end with a brief presentation of an
application to the geological modelling of the Hamersley iron ore
district, Australia, and a short discussion.
Importantly, the dual kriging scheme given in Appendix A,
gives a mathematical basis for interpolating geological
observations where the observed mapping contacts and dips/
strikes are quite sparse. The interpolation scheme for a geo-
logical series is naturally conformable, yielding realistic 3D
surfaces that are close to ‘balanced’, while following the geo-
logical trends. The geostatistical methods used to achieve this
end provide unbiased and optimal interpolation outcomes.
Geology equations
To characterise a geological series using the mathematics of
potential fields:
• each observation of a contact adds one equation to a ‘global’
matrix,
• each dip vector contributes three equations,
1
D FITZGERALD, J-P CHILÈS and A GUILLEN
• the drift for each fault contributes an equation, and
• the detrending of geology to support universal kriging adds
up to ten equations.
The extra covariance terms of the system are dominated by the
structural data. It is this that makes it possible to use a cubic
cokriging model to translate the standard deviations of the dip
directions to potential field estimates.
The system of equations for each series forms a square matrix
that is positive definite. It has been solved using Gauss
elimination. Optimisation efforts include use of a Cholesky
vector processing and principal component analysis. The degree
of smoothing of the predicted geological surfaces is directly
controlled by the range of the variogram for the series.
From the recent study by the Geological Survey of Victoria,
the regional 3D Bendigo model, an Ordovician series was
modelled using 102 structural observations (three component
vector) and 1582 contacts.
A principal component analysis shows the total domination of
the structural data terms in the interpolator as seen in Figure 1.
There are about 310 equations that are important and the rest
make only minor contributions. This demonstrates the principal
that ‘less is more’ when it comes to using geological contacts,
depending upon the required smoothness and the scale of your
project.
FIG 1 - A principal components analysis of the dual kriging equation system used to interpolate the ordovician units in the 3D Bendigo
model. The first 310 components are derived from structural observations and the rest are the geological contacts.
2 Perth, WA, 27 - 29 July 2009
PRACTICAL IMPLEMENTATION ISSUES
The potential field method has been implemented in Geo-
Modeller (www.geomodeller.com), initially developed by BRGM
(the French Geological Survey) and now commercialised by
Intrepid Geophysics. Significant support from a consortium led
by Geoscience Australia has also been shown, with the
development of an integrated stochastic, lithologically constrain-
ed geophysical inversion module, and more recently, the addition
of geothermal simulation capabilities.
In order to model real-world situations a number of practical
implementation issues had to be solved. Apart from occasional
sedimentary examples, a geological body rarely exists through-
out a domain. Geological events usually lead to complex
topology where formations cut across or onlap onto each other as
a result of deposition, erosion, intrusion or hiatus.
Such geology can be modelled by combining multiple
potential fields and the use of universal kriging principals.
Modelling several interfaces
In practical applications when several interfaces are modelled
several potential fields are then used. Overturning of the geology
due to extensive folding, faulting and other processes can be
accommodated. The method supports modelling of realistic 3D
geometries of intrusives. The important first step for the geologist
is to define a stratigraphic column. This determines how to
combine the various potential fields. The column defines the
Iron Ore Conference
 DELINEATE THREE-DIMENSIONAL IRON ORE GEOLOGY AND RESOURCE MODELS
chronological order of the interfaces as well as their nature, coded
as either ‘erode’ or ‘onlap’. For example, an ‘erode’ potential field
is used to mask the eroded part of the previous formations.
Figure 2 illustrates the rules for modelling complex geology.
Different potential field functions are used for different
geological series. These multiple potential fields are managed
using Onlap and Erode relations between series. In this example
each series comprises a single formation:
• interpolated Formation 1 (basement) and data for potential
field of Formation 2;
• formation 2 interpolated using an Onlap relation and data for
potential field of formation 3; and
• formation 3 interpolated using an Erode relation.
FIG 2 - Complex geology is modelled using different potential-field
functions for different geological series. These multiple potential
fields are managed using Onlap and Erode relations between
series. In this example each series comprises a single formation:
(A) interpolated Formation 1 (basement) and data for potential
field of Formation 2; (B) Formation 2 interpolated using an Onlap
relation and data for potential field of formation 3; and (C)
Formation 3 interpolated using an Erode.
Iron Ore Conference Perth, WA, 27 - 29 July 2009
Faults
Several methods can be envisaged to handle faults. If they delimit
blocks and the potential field is not correlated from one block to
the other, it obviously suffices to process each block separately.
Another conventional technique is to consider faults as screens.
The method used in the 3D Geological Editor is the method
proposed by Marèchal (1984) to handle faults in the 2D inter-
polation of the elevation of interfaces, where faults are entered as
external drift functions. This method requires the knowledge of the
fault planes and also of the zones of influence of the faults.
Let us start with a very simple example, a normal fault
intersecting the whole study zone and dividing it in two subzones
D and D’. That fault induces a discontinuity of the potential field,
whose amplitude is not known. Cokriging can accommodate that
discontinuity whatever its amplitude by introducing a drift
function complementing the L polynomial drift functions, for
example:
f L+1(x) = 1D(x),
or equivalently, in a symmetric form:
f L+1(x) = 1D(x) – 1D’(x).
If the polynomial drift functions include the monomial
f 1(x) = x (first coordinate) due to the presence of a linear trend of
the potential field, and we have good reasons to suspect not only
a discontinuity but also a change of slope of the drift when cross-
ing the fault, it is advisable to also introduce an additional drift
function such as f L+2(x) = x 1D(x).
A finite fault (limited extent) is modelled with a drift function
with a bounded support. The fault vanishes on the support
boundaries; inside that support, the function takes on positive
values on one side of the fault plane, with a maximum at the
centre of the fault, and negative values on the other side. Figure 3
illustrates how that method takes faults into account. In real-world
FIG 3 - Handling faults. Top – data points located on two interfaces
and structural data; middle – model built without introducing any
fault; bottom – model taking faults into account.
3
D FITZGERALD, J-P CHILÈS and A GUILLEN
applications a fault plane is not a planar surface. It is often only
known by some points on its surface and unit vectors orthogonal
to it. Its geometry is also modelled by a potential field.
Boreholes
The primary use of boreholes in this method is to provide
observations of the contacts between different lithologies. This
requires a mapping from the detailed downhole logs to the scale
at which you wish to work. Recent work has been directed at
making this much easier for the geologist. Figure 4 shows a
borehole log and a corresponding borehole section through a 3D
project and a demonstration where the misfit is less than 0.2 per
cent overall. Sometimes a fault may cross the borehole. An
ability to re-interpret the lithological log interactively and add a
fault contact can be an important means of getting the 3D
geology interpretation to work.
FIG 4 - A reconciliation of the borehole lithology log against the 3D
model. The dual kriging technology knows the data to better than
0.2 per cent, whilst also accommodating all surface mapping, etc.
CHALLENGES FOR MODELLING IRON ORE
DEPOSITS
As with all thin bed, stratigraphically controlled geological units,
the challenges for the project geologist that must be addressed in
the modelling are:
4 Perth, WA, 27 - 29 July 2009
• Develop a concept for spatial distribution of lithology.
• Transverse isotropic interpolation of the beds – an
anisotropic covariance is used to model thin beds less than
1 m thick over a lateral extent covering many kilometres.
• Vertical exaggeration during visualisation – this is important
to enable fine tuning of the economic horizons in the context
of a large lateral extent.
• Limited faults – local limited faults can be modelled easily
and modified to gauge their influence.
• Forward modelling of the gravitational response of the
geology.
Independently observed geophysical data sets are commonly
available. They provide a very important means of checking the
model. This includes an ability to model real topography and
high rock density units in limited surface relief. One aim here is
to simulate what would be observed from a low flying aircraft
with a next generation gravity gradiometer on board. The other
aim is to use ground gravity as an independent tool to check the
fit of the model to the ‘reality’.
LITHOLOGICAL PROPERTY ESTIMATION AND
MODELLING
Both a drilling database and detailed observations of topography
and gravity should be used in estimating the density of the
orebody in the model. We are involved in an on-going study with
Rio Tinto Exploration to demonstrate how sensitive an airborne
gravity gradiometer (AGG) needs to be to compete with the
accuracy and usability of ground based gravity acquisition.
Existing systems, FALCON, Bell Geospace and Arkex are
generally thought to resolve to no better than 8 Eö/√Hz or in
layman’s terms a difference in the gravitational acceleration
locally of eight parts in 109 is lost in the noise.
The setting for these tests is the Pilbara where there is
considerable topographic relief (>100 m) associated with an
unweathered near-surface iron ore deposit. This buried deposit
has a large volume and has a higher density than the surrounding
host rocks.
The airborne gravity survey has an average drape clearance of
80 m with up to ± 50 m near the cliff top. A detail digital terrain
model with a spatial resolution of better than 25 m and a vertical
resolution to ± 5 cm was used in this study. The aim is to test
how sensitive a next generation gravity gradiometer instrument,
being flown in a conventional survey aircraft, needs to be to find
and delineate iron ore resources quickly and efficiently. There
are three well advanced teams working on next generation
instruments namely, Rio Tinto, Gedex and Arkex. Our work
shows an instrument with an error of around 1 Eö/√Hz would
deliver a powerful exploration tool with significantly improved
capability of resolving near-surface density anomalies.
Figure 5a shows the acquired gravity gradient signal before
any attempt is made to remove the terrain effects. The data is
processed to continue the signal to a smoothed drape surface that
is a good approximation to the average clearance. This removes
all flight line based biases.
A classical ‘hammer’ method terrain correction is then applied
to remove the terrain effects, assuming the background rock
density is 2.67 g/cc. Figure 5b then shows the remaining density
anomaly map. In this case, the ‘target’ orebody is the one shown
in the cliff face. Other types of buried targets are present, but are
not of interest to the subject of this paper.
Without doubt, one of the greatest weaknesses in creating 3D
geological models to use in both exploration mapping and
resource estimation, is the assigning of realistic lithological
properties to the model. Geophysical surveys of gravity gradio-
metry has an important part to play here. The integration of
Iron Ore Conference
 DELINEATE THREE-DIMENSIONAL IRON ORE GEOLOGY AND RESOURCE MODELS
(A)
(B)
FIG 5 - (A) Simulation of the cross line gravity gradient terrain
response of DTM from LIDAR data, 10 m cell size; and (B) terrain
corrected gravity shows the same small escarpment now with an
embedded high density ‘iron ore’ deposit.
density and lithology to produce a detail forward model of the
predicted 3D gravity response of the mapped area is an important
check that the model is reflecting independently observed gravity
data sets to an acceptable level. Importantly you do not have to
assume homogeneity of properties as you also have a 3D geology
model to help interpret your data.
Estimating tonnes and grade
Once the various lithological units have been delimited, we have
to tackle the estimation of grade and tonnage. This can be done
with geostatistical techniques, namely kriging, or better with
cokriging in order to simultaneously and consistently estimate the
iron grade and the grade of by-products and penalty substances.
At the local scale (eg a core or a small block) the ore tonnage
is the product of ore volume by ore density; similarly, the metal
tonnage is the product of ore tonnage by ore grade (expressed in
weight percentage). In the case of an iron ore deposit there is a
high correlation between ore grade and ore density, which shall
be taken into account. If the measurements include ore density
and ore grade for all the samples, it suffices to work with the
Iron Ore Conference Perth, WA, 27 - 29 July 2009
FIG 6 - Experimental variogram is derived directly from the 3D
point data. The model variogram is then used to interpolate in 3D,
the estimated quantity.
volumetric grade, namely the product of ore density by ore grade
in percentage. Otherwise, the correlation between density and
grade shall be studied. In both cases we have to model ore
density in order to estimate the ore tonnage (See Figures 6
and 7). This shall be done on the basis of the data available at
different scales (geophysical interpretation of gravity data,
analysis of bulk samples or core samples, etc). Multivariate
geostatistics provides tools for that integration (support change
modelling, cokriging, external drift kriging, etc).
An important issue for the sound application of geostatistics is
the correct modelling of spatial correlations. In subhorizontal
deposits, the lateral grade variations are usually much smoother
than the vertical ones and the variogram analysis considers the
horizontal variogram and the vertical variogram. In more
complex layered deposits, the analysis of spatial correlations
shall consider the variations along the layers and orthogonal to
them. This is done by ‘horizontalising’ the data. The fact that the
geological model has been built with the potential field approach
provides a consistent means to perform that step. For example, if
a layer is defined by two potential values of a common potential
field, the value of the potential at any point in the layer can be
used as a new vertical coordinate. In the system defined by the
original horizontal coordinates and the new vertical coordinate,
the main anisotropy directions are the horizontal and vertical
directions, so that the analysis of spatial correlations can be
carried out in the usual way. Kriging can be done in that system
and then exported in the original physical coordinate system.
APPLICATION TO THE HAMERSLEY DISTRICT
The geological scale and purpose of the model can vary
enormously. GeoModeller has been used for regional scale
geological modelling of the Alps and the Massif Central in
Europe. For example, Maxelon (2004); Maxelon and
Mancktelow (2004), used it to model foliation fields and a
juxtaposition of nappes with a strong folding in the Lepontine
Alps.
Australian regional cases include the Gawler Craton, Bendigo,
Burdekin 3D studies.
5
D FITZGERALD, J-P CHILÈS and A GUILLEN

FIG 7 - Cut-off grade can be imposed by selecting via a histogram, the portion of the population on interest.

FIG 8 - Geological section of Brockman iron-formation hosted orebody (from unpublished internal BHP Billiton report). This together with
sparse drill hole data is used to capture geological uncertaintly in grade-tonnage estimates, using the potential method.

6 Perth, WA, 27 - 29 July 2009 Iron Ore Conference

 DELINEATE THREE-DIMENSIONAL IRON ORE GEOLOGY AND RESOURCE MODELS

At the mining scale, the Broken Hill, Guillen et al (2004),

Bendigo Gold Mine, San Nicholas, Lane (2008) and Peruvian
Andes studies indicate the diversity and complexity of the
geological environments. This method has been applied to the
Hamersley region by various groups. The recent paper by
Osterholt, Herod and Arvidson (2009) details how BHP Billiton,
in association with SRK, are routinely using the method to report
exploration target size and type with ranges of uncertainty
compliant with the Australasian Code for Reporting Mineral
Resources and Ore Reserves (the JORC Code). Figure 8 is
reproduced with permission from this paper. Data during early
evaluation work is usually sparse and historically not sufficient to
support public reporting of resources. They give a methodology
to address the uncertainty using an holistic view to develop:
• geology and grade scenarios,
• 3D geology modelling to create the volumes, and
• grade modelling.
For this paper, we report on some work done in the Hamersley
to build a 3D model using these methods, using the stratigraphic
pile shown in Figure 9. A study area 5 km × 2 km × 1 km was
chosen. The iron ore bearing formations are folded and faulted
and then overlain by colluvium or recent sediments as shown in
section ??? in Figure 10. The beds are extensive laterally. The
desire to model thin beds over an extensive area was one of the FIG 10 - Geological units and the relationships, showing the onlap
study objectives. Vertical exaggeration of up to three to one or erosional relationship between different series.
assists in this task.

FIG 11 - Plan view of the geological model. The colours

correspond to the geological units shown in Figure 10. The
presence of longitudinal faults are clearly seen. The project covers
an area 5 km × 2 km × 1 km thick

FIG 9 - Representative section of a 3D model created to model

interaction of folding and faulting on the Brockman formation.
Vertical exaggeration is set to 2:1.

Many similar sections are created and interpreted, as well as

the geology at the topographic surface. Borehole lithology data is
also used to constrain the third dimension, as each of the
formations is modelled. The 3D model is realised by calculating
each of the series independently and then applying the onlap/
erode rules to resolve the final layout. FIG 12 - Forward model of the vertical gravity component (Gz) at a
Figure 11 shows a 3D perspective view of the geological model. fixed elevation above the plan view of the iron ore geological
The colours correspond to the geological units shown in Figure 10. model. This is normally what is collected on the ground. Units are
The presence of three longitudinal faults is clearly seen. mGals.
Figure 12 shows the near surface gravity response that would
be expected from the model. There is a desire to extend the sensitivity of geophysical
This is a very useful independent check that the model and instruments to enable a better realisation of density anomalies
observed gravity are in close agreement. and the geometries of orebodies. Figure 13 shows what might be

Iron Ore Conference Perth, WA, 27 - 29 July 2009 7

D FITZGERALD, J-P CHILÈS and A GUILLEN
FIG13 - Forward model of the gravity gradient Gzz (top) and Gyz
(bottom) at a fixed elevation above the plan view of the iron ore
geological model. Units are Eotvos. This is the vertical and north
gradients of the usual gravity measurement. There is not a clear
expression of the faults in these images. The gravity gradient data
indicates the folded nature of the higher density rocks.
expected if a full tensor gravity gradiometer was used in this
area. The signature of the faults in the gravity is weak. This is
where a full tensor magnetic gradiometer system would help.
Recent extensions to the GeoModeller technology include:
• An integrated borehole, conventional geostatistical capability
as described above. This initiative is being under taken in
association with Geovariance.
• Speed and detail enhancements to the prediction of the
geophysical responses by using 3D Fast Fourier Transform
technology.
• Batch scripting for the high fidelity rendering of geological
contacts and faults.
• Predicting the temperature gradients based upon thermal
conductivity properties and heat production rates.
POTENTIAL FIELD METHOD SUMMARY
The method presented here is designed for 3D geological models
of ore deposits built from interface points and polarised
orientation data. The methodology is designed for cases where
the geology is known at sparse locations, eg when data are
available on the surface but not at depth. The orientation data, ie
dip measurements, are not necessarily located on the geological
interfaces. They can represent stratifications or foliations related
to the contacts. Data are interpolated through a potential field
implicit function continuously defined in the entire 3D domain.
Thus, the model predicts the geological formation at any 3D
point.
Geological interfaces in the model are particular isosurfaces
extracted from the potential field. They may have any kind of
3D geometry: multilayer type, recumbent folds, complex
intrusions, etc.
The geometry of faults is computed by applying the same
method. Faults can be infinite within the 3D domain, interrelated
in a fault network, or finite.
The throw of the faults are predicted from the other field
observations and do not need to be modelled in detail.
Anisotropic interpolation of thin beds allows the geologist to
control the geological sequence over many kilometres with
sparse data observations. Inequality constraints such as a
borehole finishing within granite are also handled using a
‘Gibbs’ iterative solver.
8 Perth, WA, 27 - 29 July 2009
Geological rules are defined to model complex geology where
formations onlap onto or erode another. These rules are also used
to automatically assign the right geological interface between
two consecutive formations. This methodology automatically
provides the intersections between geological units, enables fast
modelling and allows the geologist to focus on geological
interpretation.
As the geological pile defines the topology of the model, one
can modify it without changing the basic data to produce
alternative interpretations and geometries. This capability makes it
possible to progressively update the model when new data or
interpretation is available. It is this ability to quickly realise several
scenarios that has found favour with the BHP iron ore group.
Future work
GeoModeller has a very active development program. In a short
time frame it is expected that:
• inferred apparent dip of structures from a seismic section will
be supported,
• thin bodies similar to the current fault modeling will be
supported,
• faults will displace faults, predicting their throw,
• simulation of geological and geothermal uncertainties will be
formalised,
• data rich portions of the project show higher fidelity, and
• geostatistics for property, tonnes and grade can be made via a
direct ISATIS plug-in.
Another future possible extension of the fundamental approach
outlined here concerns the geological gradient. The gradient of a
random function is rarely a unit vector. GeoModeller treats the
structural data as a unit vector ignoring the ‘strength’ of the trend.
The ideal would be to sample both a structural direction and a
structural intensity, but this is possible only in very specific cases.
Aug (2004) has shown on simulations of actual situations that
replacing actual gradients by unit vectors usually has a minor
impact on the determination of the covariance and the cokriging.
A useful improvement of the method may be to extend the
interpolation to support an optional ‘strength or intensity’ value.
CONCLUSION
Both geology and geophysics practice needs re-engineering to
simplify the identification of buried economically significant
resources. The new geoscience framework includes:
• Quantitative and repeatable geology in 3D. The decisions are
‘what scale’ and ‘what purpose’.
• Airborne systems that deliver gravity and magnetic
signatures of rocks ten times more precisely than 1980
technology. The key here is driving noise from instruments
towards 1 Eotvos or 100 pico Tesla per metre (pT/m).
• Appropriately built 3D geophysical simulation models from
the geology to help create the ‘right’ interpretations.
This sensible joining of the disciplines of structural geology
interpretation, resource estimation and computational geophysics
provides a novel method for increasing the productivity of senior
geoscientists leading to faster and better 3D modelling of
orebodies. The integration of gravity and gradiometry provides
independent checking for the model and helps to constrain the
economic geology.
The rapid delineation of the iron ore resources, using an
implicit lithology model based upon all mapping and sparse
drilling provides estimates that are much closer to the JORC
(Joint Ore Reserves Committee) code spirit than just using
polygonal based estimates.
Iron Ore Conference
 DELINEATE THREE-DIMENSIONAL IRON ORE GEOLOGY AND RESOURCE MODELS

ACKNOWLEDGEMENTS • Structural data – in the case of sedimentary rocks the

stratification is parallel to the geological horizons. We
Theo Aravanis of Rio Tinto Exploration initiated work on measure a unit vector normal to the stratification. They can
sensitivity studies for detecting buried iron ore deposits using also be unit vectors orthogonal to foliation planes for
gravity gradiometry. metamorphic rocks. Measurements are made on outcrops or
BHP Iron Ore kindly allowed the inclusion of their exploration in boreholes, either on the interfaces or anywhere within a
resource model. formation.
Geological Survey of Victoria has allowed the mention of the
Bendigo 3D geological model discussion.
The Australian Government has funded Intrepid Geophysics
via a Commercial Ready Grant.
The 3D FFT work was stimulated by Jeff Phillips of the USGS.
The geostatistical research work carried out at the École des
Mines de Paris was funded by BRGM.

REFERENCES
Aug, C, 2004. Modélisation géologique 3D et caractérisation des
incertitudes par la méthode du champ de potentiel, PhD thesis, École
des Mines de Paris, 198 p.
Calcagno, P, Chilès, J P, Courrioux, G, Guillen, A, 2008. Geological
modelling from field data and geological knowledge: part 1,
modelling method coupling 3D potential field interpolation and
geological rules, Physics of the Earth and Planetary Interiors,
171:147-157.
Chilès, J P, Delfiner, P, 1999. Geostatistics: Modelling Spatial
Uncertainty, 695 p (Wiley: New York).
Lane, R, McInerney, P and Seikel, R, 2009. Using a 3D geological
mapping framework to integrate AEM, gravity and magnetic
modelling – San Nicolas case history, in Proceedings Australian
Society of Exploration Geophysicists 18th Geophysical Conference
and Exhibition.
Mallet, J L, 2003. Geomodelling, 599 p (Oxford University Press: Oxford).
Maréchal, A, 1984. Kriging seismic data in presence of faults, in
Geostatistics for Natural Resources Characterization (eds: G Verly,
M David, A G Journel and A Maréchal), part 1, 271-294 (Reidel:
Dordrecht).
McInerney, P, Golberg, A and Holand, D, 2007. Using airborne gravity data
to better define the 3D limestone distribution at the Bwata Gas Field, FIG A1 - Principle of the potential-field method. Top – surface
Papua New Guinea, in Proceedings Australian Society of Exploration data-points at interfaces and structural data; bottom – vertical
Geophysicists 18th Geophysical Conference and Exhibition, Perth. cross-section through the 3D model.
Osterholt, V, Herod, O and Arvidson, H, 2009. Regional three-dimensional
modelling of iron ore exploration targets, in Proceedings Orebody
Modelling and Strategic Mine Planning, (ed: R Dimitrakopoulos) pp For the interpolation of the potential field, these data are coded
35-42 (The Australiasian Institute of Mining and Metallurgy: as follows:
Melbourne).
• Since the potential value at m + 1 points x0, x1, …, xn
APPENDIX A – BASIC PRINCIPLE OF THE sampled on the same interface is not known, these data are
POTENTIAL FIELD METHOD taken as m increments T(xα) – T(x’α), α = 1, …, m, all valued
to zero. Two classical choices for x’ consist in taking either
The basic method is designed to model a geological interface or a the point x0 whatever α, or the point x α–1 (the choice has no
series of subparallel interfaces Ik, k = 1, 2, … (Calcagno et al,
impact on the result; other choices are possible provided that
2008). The principle is to represent the geology by a potential
the increments are linearly independent). Since the sampled
field, namely a scalar function T(x) of any point x = (x, y, z) in 3D
data can be located on several interfaces, let M represent the
space, designed so that the interface Ik corresponds to an iso-
potential surface, ie the set of points x that satisfies T(x) = tk for total number of increments (it is equal to the total number of
some unknown value tk of the potential field. Equivalently, the data points on the interfaces minus the number of interfaces).
geological formation encompassed between two successive • The unit vector normal to each structural plane is considered
interfaces Ik and Ik’ is defined by all the points x whose potential as the gradient of the potential field, or equivalently as a set
field value lies in the interval defined by tk and tk’. In figurative of three partial derivatives •T(x) / •u, •T(x) / •v, •T(x) / •w at
terms, in the case of sedimentary deposits T could be seen as the some point xβ. The coordinates u, v, w are defined in an
time of deposition of the grain located at x, or at least as a mono- orthonormal system; this system can be the same for all the
tonous function of that geological time and an interface as an points or a specific system can be attached to each point (the
isochron surface. result does not depend on the choice provided that the three
partial derivatives are taken in consideration). In the sequel
Data types let •T(x ) /•u denote any partial derivative at x and N
denote the total number of such data (in practice N is a
T(x) is modelled with two kinds of data, as shown in Figure A1: multiple of 3 and the x form triplets of common points). Let
• Points known to belong to the interfaces I1, I2, …, typically us recall that the x do not necessarily coincide with the x
3D points discretising geological contours on the geological (the latter are located on the interfaces whereas the former
map and intersections of boreholes with these interfaces. can be located anywhere).

Iron Ore Conference Perth, WA, 27 - 29 July 2009 9

D FITZGERALD, J-P CHILÈS and A GUILLEN
Interpolation of the potential field
The potential field is then only known by discrete or
infinitesimal increments. It is thus defined up to an arbitrary
constant. So an arbitrary origin x0 is fixed and at any point x the
potential increment T(x) – T(x0) is kriged. The estimator is in fact
a cokriging of the form:
Μ
N
∂T
T (x) − T (x 0 ) = ∑ μ α ( T(x α ) − T(x′ α )) + ∑ ν β
∗ ∗
(x )
∂uβ β
α =1 β =1
where the weights µα and vβ, solution of the cokriging system,
are in fact functions of x (and x0). One may wonder why the
potential increments are introduced in that estimator since their
contribution is nil. The key reason is the weights are different
from weights based on the gradient data alone. Conversely, the
gradient data also play a key role, because in their absence the
estimator would be zero for any x.
Cokriging is performed in the framework of a random function
model. T(.) is assumed to be a random function with a
polynomial drift:
L be considered as an upper bound for C when the potential
m(x) = ∑ b l f l (x)
l=0
and a stationary covariance K(h). Since the vertical usually
plays a special role, the degree of the polynomial drift can be
higher vertically than horizontally and the covariance can be
anisotropic. For example, if we model several subparallel and
subhorizontal interfaces, it makes sense to assume a vertical linear
drift of the form m(x) = b0 + b1 z, ie with two basic drift functions
f 0(x) • 1 and f 1(x) = z. A geological body with the shape of an
ellipsoid would correspond to a quadratic drift, ie to the ten basic
polynomial coefficients with degree less than or equal to two.
Once the basic functions f •(x) of the drift and the covariance
K(h) of T(.) are known, we have all the ingredients to perform a
cokriging in the presence of gradient data, as shown in ChilPs and
Delfiner (1999, section 5.5.2). Indeed, the drift of •T(x) / •u is
simply Mm(x) / Mu, ie a linear combination of the partial
derivatives •f •(x) / •u with the same unknown coefficients b. as for
m(x), the covariances of partial derivatives are second-order partial
derivatives of K(.), and the cross-covariances of the potential field
and partial derivatives are partial derivatives of K(.).
Implementation of the cokriging algorithm
Since the potential increment data in fact do not contribute to the
final cokriging estimate, the estimator can be seen as an
integration of the gradient data. To preserve the spatial continuity
of the cokriging estimates it is wise to work in a unique
neighbourhood, namely to effectively include all the data in the
cokriging of T(x) for every x. If we are not interested in the
cokriging variance, cokriging can be implemented in its dual form,
which has two advantages:
1. the cokriging system is solved once; and
2. that form is especially suited when cokriging is considered
as an interpolator, because it allows an easy estimation of
T(x) - T(x0) at any new point x.
The latter property is very useful to display 3D views of the
geological model with an algorithm such as the marching cube,
which starts from the estimation of T(x) - T(x0) at the nodes of a
coarse regular grid and then requires intermediate points to be
predicted to track the desired isopotential surface.
INFERENCE OF THE COVARIANCE OF THE
POTENTIAL FIELD
In usual geostatistical applications, the covariance or variogram
of the variable under study is modelled from the sample vario-
10 Perth, WA, 27 - 29 July 2009
gram of the data. In the present case, we have few measurements
of the potential T(x), and the potential increments used for the
interpolation cannot be used for the inference of K since they all
have a zero value. The choice of the model followed from these
considerations:
• At the scale considered, geological interfaces are smooth
rather than fractal surfaces which implies that the covariance
is twice differentiable. A cubic model was considered a good
compromise among the various possible models, because it
has the necessary regularity at the origin and a scale parameter
(the range) which can accommodate various situations.
• The scale parameter a and sill C of the covariance K(h)
determine the sill of the variogram of the partial derivatives:
it is equal to 14C/a2 in the case of an isotropic cubic
covariance considered here. When there is no drift and the
geological body is isotropic (eg a granitic intrusion), the unit
gradient vector can have any direction so that its variance is
equal to one. The variance of each partial derivative is then
equal to 1/3. A consistent choice for C once the scale para-
meter a has been chosen is thus C = a2/42. That value shall
field has a drift, because in that case the mean of the
potential gradient is not equal to zero so that its variance is
shorter than one (its quadratic mean is one by definition).
• Sensible measurement variances can also be defined (nugget
effects).
The assumption of an isotropic covariance model is too
restrictive and can be relaxed. In practice the covariance K(h) is
supposed to be the sum of several cubic components Kp(h), each
one possibly displaying a zonal or geometric anisotropy. To
avoid too much complexity, the main anisotropy axes u, v, w, are
common to all the components of a series.
Thanks to these formulae the covariance parameters of K
(nugget effect, scale parameter of each covariance component in
the three main directions, sill of each component) are chosen so
as to lead to a satisfactory global fit of the directional sample
variograms of the three components of the gradient. An
automatic fitting procedure based on the Levenberg-Marquardt
method has been developed to facilitate that task (August 2004).
Figure A2 shows an example of such a fitting. One thousand,
four hundred and 85 structural data were sampled in an area of
about 70 × 70 km2 in the Limousin (Massif Central, France). The
main (u, v, w) coordinates here coincide with the geographical
(x, y, z) coordinates. Since the structural data are all located on
the topographic surface, the variograms have been computed in
the horizontal plane only. Note that the sill of the variogram of
the vertical component is much lower than that of the horizontal
components. This is due to the fact that the layers are
subhorizontal so that the vertical component of the gradient
displays limited variations around its non-zero mean. The model
K includes three components, the second of which only depends
on the horizontal component of h and the third one on the
north-south component (zonal anisotropies).
UNCERTAINTY ON THE THREE-DIMENSIONAL
MODEL
Case studies have shown that the use of a sound covariance
model improves the model in comparison with the use of a
conventional model. An additional interest in using a covariance
fitted from the data is the possibility of obtaining sensible
cokriging standard deviations.
When the ‘true’ covariance of the potential field is known, a
meaningful cokriging standard deviation CK(x) can be associated
with the cokriging of T(x) - T(x0). The calculation of that
standard deviation requires the use of the standard form of the
cokriging system, which calls for more computing time than its
dual form (this is the price to pay for knowing the uncertainty
Iron Ore Conference
 DELINEATE THREE-DIMENSIONAL IRON ORE GEOLOGY AND RESOURCE MODELS

0.3 assumption which seems reasonable in the present context, the

probability that a given point x belongs to that formation is:
γY//
( )⎞⎟
⎛ t' − T ∗ (x) − T ∗ (x 0 )
γX// γY ⊥
{
Pr t ≤ T(x) − T x 0( ) } < t' = G⎜
⎜ σ CK (x) ⎟
0.2 ⎝ ⎠
Variogram

where:
γX ⊥

G is the standard normal cumulative distribution function.

0.1
Similarly, if we are interested in the interface passing through
the point x0, namely in the set of points x such that
T(x) - T(x0) = 0, the variable R(x) = [T*(x) - T*(x0)]°/ CK measures
γZ
⊥
the likelihood that x belongs to the interface. Indeed, writing the
obvious relation:
0.0 T(x) - T(x0) = T*(x) - T*(x0) + cokriging error,
0 5000 10 000 15 000 20 000 25 000
we see that x belongs to the interface if and only if
Distance T*(x) - T*(x0) is equal to minus the cokriging error, or
equivalently if R(x) is equal to minus the standardised cokriging
FIG A2 - Example of fitting of the covariance of the potential field error (the ratio of the error by CK(x)). The value of that error is
from the sample variograms of the partial derivatives of the not known but it is a variable with zero mean and unit variance.
potential field. Limousin data set, Massif Central, France. γX// and For example, assuming again that the potential field is
γX denote the variogram of the partial derivative •T/•x Gaussian, the area defined by |R(x)| <2 includes about 95 per
cent of the actual interface. Figure A3 displays R(x) for the top
of the lower gneiss unit in the Limousin.
The black line corresponds to R(x) = 0, ie to the isovalue
surface of the cokriged potential field passing by the data points
sampled on that interface. The true interface is likely to be found
in the light coloured area, whereas the darkest area can be
considered as a forbidden area. This capability is not routinely
made available within GeoModeller.

FIG A3 - Representation of the uncertainty of the top of a

geological unit by the variable R(x) (upper gneiss unit, Limousin).
The data (geological map and structural data) are all located on
the topography. Top – map of a zone of 65 km × 65 km in the
horizontal plane with elevation 500 m; bottom – vertical eastwest
cross-section with 62 km extension and 34 km depth. The black
curve represents the kriged interface. The true interface is in fact in
the coloured zones, with a smaller probability as the zone is darker.
The darkest zones can be considered as exclusion zones (after
August 2004).

attached to the geological model). Let us suppose that some

geological formation is defined by the set of points x such that
T(x) - T(x0) is comprised between two values t and t’. Assuming
that the potential field is a Gaussian random function, an

Iron Ore Conference Perth, WA, 27 - 29 July 2009 11

12 Perth, WA, 27 - 29 July 2009 Iron Ore Conference

View publication stats