Mining Goes Digital – Mueller et al. (Eds)
© 2019 Taylor & Francis Group, London, ISBN 978-0-367-33604-2
210
Transforming exploration data through machine learning
I.W.S. Whitehouse
Geobank, Australia
W. Slabik
MICROMINE, Australia
ABSTRACT: The application of machine learning to the process of collecting and
analysing geological data in mineral exploration has the potential to transform the way
explorers operate.
The traditional process of plan – drill – observe – measure – analyse can be slow and lead
to costly re-drill or re-sample programs. A common issue faced by exploration companies is
the inconsistency in the way data has been collected and categorised. This complicates the
task of data modelling when undertaking economic viability studies. Using machine learning,
data can be cleansed and validated prior to starting the modelling process.
There are several ways to streamline the process for the resource geologist, the first
being feature identification through imagery. High quality DSLR cameras provide a tool
for exploration companies to collect high quality imagery of core and chip trays. Machine
learning algorithms can recognize features in the images such as colour, structures, veins,
particle size and hand-written text. It is feasible for this data to be automatically collected
and stored in a database.
Drill hole databases record rock interval attributes like rock code, hardness, colour, grade,
location, and geophysical measurements. These attributes could be used as a lithological
signature to identify other instances of similar signatures within the database. This technique
could be used for data consistency testing or to discover new information within the dataset.
Finally, to illustrate the power of machine learning, a small research project is presented
that successfully identified the regions of core tray imagery that contained drill core.
1 INTRODUCTION
Machine Learning (ML) is the field of study that gives computers ability to learn without
being explicitly programmed (Géron, 2017). Instead of designing computer programs with
explicit complex logic, the programs use input data to self-learn patterns to generate the
desired output. The beginnings of machine learning and artificial intelligence (which ML
is part of) date back to the early 1950s when Marvin Minsky and Dean Edmonds built the
first computer machine capable of learning, (http://cyberneticzoo.com/mazesolvers/1951maze-solver-minsky-edmonds-american/). However, technological limitations made it
impossible to design and build a practical system that could help with complex problems.
In 1996, when IBM developed a chess playing ML system that beat the reigning world
champion (https://www.ibm.com/blogs/think/2017/05/deep-blue/), a significant development milestone in ML was achieved. This was followed by the introduction of AlexNet
– machine learning neural network (Krizhevsky, A., Sutsskever, I., Hinton, G. 2012) which
won the 2012 ImageNet competition in which a program had to label 150,000 photographs
(http://image-net.org/challenges/LSVRC/2012/). AlexNet proved that computers can learn
difficult patterns. These developments in ML design and massive advancements in computer processing speeds since the 1950s, means ML solutions are now able to accurately
process information in real time.
© 2019 Taylor & Francis Group, London, ISBN 978-0-367-33604-2
210
Transforming exploration data through machine learning
I.W.S. Whitehouse
Geobank, Australia
W. Slabik
MICROMINE, Australia
ABSTRACT: The application of machine learning to the process of collecting and
analysing geological data in mineral exploration has the potential to transform the way
explorers operate.
The traditional process of plan – drill – observe – measure – analyse can be slow and lead
to costly re-drill or re-sample programs. A common issue faced by exploration companies is
the inconsistency in the way data has been collected and categorised. This complicates the
task of data modelling when undertaking economic viability studies. Using machine learning,
data can be cleansed and validated prior to starting the modelling process.
There are several ways to streamline the process for the resource geologist, the first
being feature identification through imagery. High quality DSLR cameras provide a tool
for exploration companies to collect high quality imagery of core and chip trays. Machine
learning algorithms can recognize features in the images such as colour, structures, veins,
particle size and hand-written text. It is feasible for this data to be automatically collected
and stored in a database.
Drill hole databases record rock interval attributes like rock code, hardness, colour, grade,
location, and geophysical measurements. These attributes could be used as a lithological
signature to identify other instances of similar signatures within the database. This technique
could be used for data consistency testing or to discover new information within the dataset.
Finally, to illustrate the power of machine learning, a small research project is presented
that successfully identified the regions of core tray imagery that contained drill core.
1 INTRODUCTION
Machine Learning (ML) is the field of study that gives computers ability to learn without
being explicitly programmed (Géron, 2017). Instead of designing computer programs with
explicit complex logic, the programs use input data to self-learn patterns to generate the
desired output. The beginnings of machine learning and artificial intelligence (which ML
is part of) date back to the early 1950s when Marvin Minsky and Dean Edmonds built the
first computer machine capable of learning, (http://cyberneticzoo.com/mazesolvers/1951maze-solver-minsky-edmonds-american/). However, technological limitations made it
impossible to design and build a practical system that could help with complex problems.
In 1996, when IBM developed a chess playing ML system that beat the reigning world
champion (https://www.ibm.com/blogs/think/2017/05/deep-blue/), a significant development milestone in ML was achieved. This was followed by the introduction of AlexNet
– machine learning neural network (Krizhevsky, A., Sutsskever, I., Hinton, G. 2012) which
won the 2012 ImageNet competition in which a program had to label 150,000 photographs
(http://image-net.org/challenges/LSVRC/2012/). AlexNet proved that computers can learn
difficult patterns. These developments in ML design and massive advancements in computer processing speeds since the 1950s, means ML solutions are now able to accurately
process information in real time.