Deep Learning & Artificial Intelligence in Mining -- Mining presents unique challenges to data scientists. They must walk a thin line between preserving scientific objectivity and extracting the greatest benefit from data. Learn why data scientists and geoscientists must collaborate to achieve meaningful results.
1. A MINER’S DRIFT
A Journal of Occasional Explorations
2018-MAR-27
Deep Learning & Artificial Intelligence in Mining
It starts with the data
Volume 01
Issue 01
Deep Learning & Artificial Intelligence in Mining Series
Joseph Starwood – Digital Advisor, Geologist, & Geophysicist
2. A Miner’s Drift
Page 1
Deep Learning & Artificial
Intelligence in Mining
I T S TA R T S W I T H T H E D A TA
Mining presents unique challenges to data scientists. They must walk a thin line between
preserving scientific objectivity and extracting the greatest benefit from data. However, in
specific situations, data scientists must understand the source data including its structure
and the transformations performed upon it. Without that understanding, artificial
intelligence efforts yield poor or even meaningless results.
Mining professionals rely on a staggering volume and variety of data. From exploration through extraction
and beneficiation to final shipment, miners utilize devices, systems, and applications that produce a wide
range of data. This data may include geologic observations, geochemical results, geophysical measurements,
engineering criteria, assay results, mine block models, mine planning details, productions yields, financials, and
more.
This assortment of data presents the usual problems to data scientists. These include vendor proprietary data
schemas and exchange formats, custom developed data schemas and exchange formats, structural and
semantic differences across schemas and formats, missing data values, and so on. Through experience, data
scientists are well-prepared to deal with these problems.
GEOSCIENTIFIC DATA
Geoscientific data, however, presents a unique challenge. Geologists, geochemists, and geophysicists observe
and measure many different rock types, alteration types, minerals, and elements.
To accommodate this, they utilize applications and databases that have a very high degree of data field
reuse. In other words, a given field may mean one thing in the first record and another thing in the second
record.
To the geoscientist, this is very familiar and convenient. But, for the data scientist, it is particularly vexing.
THE SOURCE DATA PROBLEM
Let’s look at this more closely. Say you are creating a new artificial intelligence (AI) solution to help a mining
company find better exploratory drilling targets.
Suppose that the core log database records the lithology, the three most prevalent alteration types with their
intensities, and the five most prevalent elements and their values among many other data items.
The data record fields would look something like this:
3. A Miner’s Drift
Page 2
Core ID, . . . Lithology, AltType1, AltType1-Int, AltType2, AltType2-Int, AltType3, AltType3-Int, Elem1, Elem1Value, Elem2, Elem2Value, Elem3,
Elem3Value, Elem4, Elem4Value, Elem5, Elem5Value, . . .
In the first record, the lithology is Greywacke. The most prevalent alteration is Tourmaline, which has a low
intensity. Gold is the most prevalent element at 20 ppm. The remaining four elements in order are Silicon,
Sulphur, Arsenic, and Iron.
In the next record, we have Schist. This time, the most prevalent alteration is Silica, which has a high intensity.
The second most prevalent alteration is Chloritization which has a moderate intensity. Silicon is the most
prevalent at 35%, Chlorine, Sulphur, Iron, and Copper follow.
That AltType1 is ‘Tourmaline’ and AltType1Int is ‘low’ in the first record while AltType1 is ‘Silica’ and
AltType1Int is ‘high’ in the second record presents a real problem.
Comparing the raw source data in such records is like comparing apples to oranges. This data is not ready
for use in an artificial intelligence or advanced analytics solution.
FIXING THE PROBLEM
Ideally, records should have fields for each alteration type and each element in which the respective values
are stored. This logic applies to any reused data field in the record.
To fix this source data problem, each data value must be read and then written to a new data record such
that the value goes into the field intended for that specific data. This process is referred to as ‘flattening’ the
data. The new data record is referred to as ‘flattened’ data.
The data record fields would look something like this:
Core ID, . . . Lithology, ChorAlt, . . . EpidAlt, OxidAlt, PyriAlt, SerpAlt, SiliAlt, . . . TourAlt, . . .Ac, Ag, Al, Am, Ar, As, At, Au, . . .Yb, Zn, Zr . . .
There are 94 naturally occurring elements. A given mining environment may exhibit a dozen alteration types.
As you can see, flattening the data results in each data record having perhaps 200+ fields.
Deciphering the source data records is not easy. Mining and geosciences have domain-specific languages
and notations. Partnering with a geologist or geotechnician helps the data scientist accelerate the data
flattening and ensure correctness.
BETTER ARTIFICIAL INTELLIGENCE
Results from several mining projects around the world demonstrate the differences between raw and flattened
data on artificial intelligence outcomes.
The typical artificial intelligence solution utilized Microsoft Azure including Machine Learning, Blob Storage,
SQL Database, and Power BI.
The precision and accuracy for predicting high-grade ore were in the mid-70 percent range for the raw (non-
flattened) data. Here the non-reused fields appear to have played a role. For the flattened data, the
precision and accuracy for predicting high-grade ore were in the high-80 percent range.
4. A Miner’s Drift
Page 3
CONCLUSION
Artificial intelligence can play a powerful role in mining. The results depend upon the completeness and
correctness of the input data. In many cases, the raw data must first be transformed to make in amenable to
deep learning and artificial intelligence. Data scientists and geoscientists must collaborate to correctly
transform the data and achieve meaningful results.
5. A Miner’s Drift
Page 4
ABOUT THE AUTHOR
Known within Microsoft Corporation as 'The Mining Guy', Joseph
Starwood is a Digital Advisor who helps mining companies, equipment
suppliers, and service providers extract value from their information
technology (IT) and operational technology (OT) investments and assets.
He co-authored the Mining Book of Dreams as well as the Microsoft
Upstream Reference Architecture for Mining (MURA-Mining).
During his early career, Joseph worked as an exploration geologist and
geophysicist. He focused on gold, silver, copper, and molybdenum
deposits in the western United States. His work included Kuroko style
gold systems. He also worked in various capacities for coal mining, oil &
gas, and heavy construction companies.
Joseph has presented at:
▪ InfraGard Great Lakes Regional Conference – 2014
▪ Society for Mining, Metallurgy, and Exploration Annual Meeting – 2014
▪ Great Lakes Software Excellence Conference (GLSEC) – 2011 & 2013
▪ Calvin College Colloquium Series – 2010
▪ Integrate: Uniting the World of IT Symposium – 2010
▪ SEI Architecture Technology User Network (SATURN) – 2009
▪ Financial Services Solutions Symposium (FSSS) – 2008 & 2009
He also contributed to:
▪ Software Engineering Institute (SEI) – Hard Problems in SOA Workshop – 2008
Contact the Author:
▪ LinkedIn: www.linkedin.com/in/josephstarwood
▪ e-Mail: JosephStarwood@JosephStarwood.com
PLEASE NOTE: The author is an employee of Microsoft Corporation. The views, thoughts, and opinions expressed in this text belong
wholly to the author, and do not necessarily reflect those of anyone else (the author’s employer, or any other group or individual).
PLEASE NOTE: All product names, logos, and brands are property of their respective owners. All company, product and service names
used in this text are for identification purposes only. Use of these names, logos, and brands does not imply endorsement.
PLEASE NOTE: The information in this text is provided without representations or warranties, express or implied.
Copyright 2018 All rights reserved
# # #