AI has caused disruption in virtually every commercial sector, as well as our day-to-day lives. However, it continues to be an enigmatic topic for the geosciences community, and how it can be harnessed remains somewhat unclear. Geoscientists stand to benefit significantly from the application of AI in their work. For example, applications like the automatic digitization of analogue data, or the use of predictive models to identify areas of mineralization are already being used successfully by some tech-forward companies. This presentation will demystify the capabilities of a key AI technology, neural networks, discuss several ways it is being used by geologists today, and outline how IBM believes advancements in this technology will continue to improve the way geoscientists work in the years to come.

2. 1. THE TECHNOLOGIES
2. Neural Networks
3. PRACTICAL APPLICATIONS
4. Current & Future Applications
5. FINAL THOUGHTS
Max Howarth - IBM Canada 2018
PROVIDE A PRACTICAL
OVERVIEW OF THE
MOST COMMON AI
TECHNOLOGY
PURPOSE:

3. Max Howarth - IBM Canada 2018

4. Max Howarth - IBM Canada 2018

5. Max Howarth - IBM Canada 2018

6. Max Howarth - IBM Canada 2018

7. 1. THE CLASSIFICATION PROBLEM
2. REGRESSION-LIKE PROBLEM

8. NEURAL
NETWORK
1. THE CLASSIFICATION PROBLEM
2. REGRESSION-LIKE PROBLEM
IT’S A CAT

9. NEURAL
NETWORK
1. THE CLASSIFICATION PROBLEM
2. REGRESSION-LIKE PROBLEM
THERE’S 30
STRUCTURAL
MEASUREMENTS

10. NEURAL
NETWORK
1. THE CLASSIFICATION PROBLEM
2. REGRESSION-LIKE PROBLEM
YOU’LL SELL 70
HOT DOGS
Date Temperature P.O.P Foot Traffic
21/09/2018 23 20% 100
22/09/2019 27 10% 321
23/09/2020 19 60% 125

11. NEURAL
NETWORK
1. THE CLASSIFICATION PROBLEM
2. REGRESSION-LIKE PROBLEM
14 DAY SALINITY
FORECAST:
680 uS/cm
Date
FLOW @
ST.1
FLOW @
ST.2
FLOW @
ST.3
1/09/2018 30 45 12
2/09/2019 22 43 15
3/09/2020 19 41 18

12. Max Howarth - IBM Canada 2018
LOTS OF EMPIRICAL DATA
NOT A LOT OF WAYS TO TIE IT TOGETHER

13. Max Howarth - IBM Canada 2018
THE GOOD
• Great for:
• Pattern recognition
• Nonlinear modelling
• Classification
• Association
• Control
• Fast
• Don’t have to explicitly
define an equation
• Learn from the past,
adapt in the future

14. THE GOOD THE BAD
• Great for:
• Pattern recognition
• Nonlinear modelling
• Classification
• Association
• Control
• Don’t have to explicitly
define an equation
• Fast
• Learn from the past,
adapt in the future
• Viewed as a panacea - but
isn’t
• Often highly situational
• Often requires intense
compute resources
• Can be hard to evaluate
success

15. Max Howarth - IBM Canada 2018
THE GOOD THE BAD THE UGLY
• Great for:
• Pattern recognition
• Nonlinear modelling
• Classification
• Association
• Control
• Don’t have to explicitly
define an equation
• Fast
• Learn from the past,
adapt in the future
• Intense data cleansing
and preparation
requirements
• “Grey box” solution
• Lack of standardized
approaches
• Hockey stick learning
curve
• Viewed as a panacea - but
isn’t
• Often highly situational
• Often requires intense
compute resources
• Can be hard to evaluate
success

16. INPUTS NEURON OUTPUT
W1
W2
W3
X1
X2
X3
f(
∑
Xn × Wn)

17. W1
W2
W3
X1
X2
X3
f(
∑
Xn × Wn)
Training
Inputs Output
0 0 1 0
1 1 1 1
1 0 1 1
0 1 1 0
1 0 0 1
TRAIN
TEST
CALCULATOR

18. W1
W2
W3
X1
X2
X3
f(
∑
Xn × Wn)
Inputs Output
0 0 1 0
1 1 1 1
1 0 1 1
0 1 1 0
1 0 0 1
TRAIN
TEST
Training
1. Set random initial starting weights CALCULATOR
W1 W2 W3
0.8 -0.2 0.7

19. W1
W2
W3
X1
X2
X3
f(
∑
Xn × Wn)
Inputs Output
0 0 1 0
1 1 1 1
1 0 1 1
0 1 1 0
1 0 0 1
TRAIN
TEST
Training
1. Set random initial starting weights
2. Calculate output using a training set
W1 W2 W3 Output
0.8 -0.2 0.7 0.7

20. W1
W2
W3
X1
X2
X3
f(
∑
Xn × Wn)
Inputs Output
0 0 1 0
1 1 1 1
1 0 1 1
0 1 1 0
1 0 0 1
TRAIN
TEST
Training
1. Set random initial starting weights
2. Calculate output using a training set
3. Quantify error between calculated output
and expected output
W1 W2 W3 Output
0.8 -0.2 0.7 0.7
Error: -0.7
Error Weighted Derivative: -0.14814

21. W1
W2
W3
X1
X2
X3
f(
∑
Xn × Wn)
Inputs Output
0 0 1 0
1 1 1 1
1 0 1 1
0 1 1 0
1 0 0 1
TRAIN
TEST
Training
1. Set random initial starting weights
2. Calculate output using a training set
3. Quantify error between calculated output
and expected output
4. Adjust training weights based on EWD
times initial input value.
W1 W2 W3 Output
0.8 -0.2 0.7 0.7
Error: -0.7
Error Weighted Derivative: -0.14814
W1 W2 W3
0.8 -0.2 0.551

22. W1
W2
W3
X1
X2
X3
f(
∑
Xn × Wn)
Inputs Output
0 0 1 0
1 1 1 1
1 0 1 1
0 1 1 0
1 0 0 1
TRAIN
TEST
Training
1. Set random initial starting weights
2. Calculate output using a training set
3. Quantify error between calculated output
and expected output
4. Adjust training weights based on EWD
times initial input value.
5. Repeat
W1 W2 W3 Output
0.8 -0.2 0.7 0.7
Error: -0.7
Error Weighted Derivative: -0.14814
W1 W2 W3
0.8 -0.2 0.551

23. W1
W2
W3
X1
X2
X3
f(
∑
Xn × Wn)
Inputs Output
0 0 1 0
1 1 1 1
1 0 1 1
0 1 1 0
1 0 0 1
TRAIN
TEST
Training
1. Set random initial starting weights
2. Calculate output using a training set
3. Quantify error between calculated output
and expected output
4. Adjust training weights based on EWD
times initial input value.
5. Repeat
W1 W2 W3 Output
0.8 -0.2 0.7 0.7
Error: -0.7
Error Weighted Derivative: -0.14814
W1 W2 W3
0.85 -0.15 0.56

24. W1
W2
W3
X1
X2
X3
f(
∑
Xn × Wn)
Inputs Output
0 0 1 0
1 1 1 1
1 0 1 1
0 1 1 0
1 0 0 1
TRAIN
TEST
Training
1. Set random initial starting weights
2. Calculate output using a training set
3. Quantify error between calculated output
and expected output
4. Adjust training weights based on EWD
times initial input value.
5. Repeat
W1 W2 W3 Output
0.8 -0.2 0.7 0.7
Error: -0.7
Error Weighted Derivative: -0.14814
W1 W2 W3
0.99 0 0
X100

25. W1
W2
W3
X1
X2
X3
f(
∑
Xn × Wn)
Inputs Output
0 0 1 0
1 1 1 1
1 0 1 1
0 1 1 0
1 0 0 1
TRAIN
TEST
Training
1. Set random initial starting weights
2. Calculate output using a training set
3. Quantify error between calculated output
and expected output
4. Adjust training weights based on EWD
times initial input value.
5. Repeat
W1 W2 W3 Output
0.8 -0.2 0.7 0.7
Error: -0.7
Error Weighted Derivative: -0.14814
W1 W2 W3
0.99 0 0
OPTIMIZE

26. W1
W2
W3
X1
X2
X3
f(
∑
Xn × Wn)
ibm.biz/sabcs2018
Try it
out

27. Max Howarth - IBM Canada 2018

28. Max Howarth - IBM Canada 2018
Groundwater Model Approximation with Artificial
Neural Network for Selecting Optimum Pumping
Strategy for Plume Removal
Shreedhar Maskey, Yonas B. Dibike, Andreja Jonoski and Dimitri
Solomatine
What is the optimal pumping rate to
minimize time to cleanup?
P1 P2 P3
40 20 30

29. Max Howarth - IBM Canada 2018
Groundwater Model Approximation with Artificial
Neural Network for Selecting Optimum Pumping
Strategy for Plume Removal
Shreedhar Maskey, Yonas B. Dibike, Andreja Jonoski and Dimitri
Solomatine
What is the optimal pumping rate to
minimize time to cleanup?
P1 P2 P3
40 20 30
MODFLOW/MODPATH
G.O. TECHNIQUE
MINIMUM TIME: 3000 DAYS

30. Max Howarth - IBM Canada 2018
Groundwater Model Approximation with Artificial
Neural Network for Selecting Optimum Pumping
Strategy for Plume Removal
Shreedhar Maskey, Yonas B. Dibike, Andreja Jonoski and Dimitri
Solomatine
What is the optimal pumping rate to
minimize time to cleanup?
P1 P2 P3
40 20 30
MODFLOW/MODPATH
G.O. TECHNIQUE
MINIMUM TIME: 3000 DAYS
MINUTES? HOURS?
THOUSANDS? MILLIONS?

31. Max Howarth - IBM Canada 2018
Groundwater Model Approximation with Artificial
Neural Network for Selecting Optimum Pumping
Strategy for Plume Removal
Shreedhar Maskey, Yonas B. Dibike, Andreja Jonoski and Dimitri
Solomatine
What is the optimal pumping rate to
minimize time to cleanup?
P1 P2 P3
40 20 30
NEURAL NETWORK
G.O. TECHNIQUE
MINIMUM TIME: 3000 DAYS

32. Max Howarth - IBM Canada 2018
Groundwater Model Approximation with Artificial
Neural Network for Selecting Optimum Pumping
Strategy for Plume Removal
Shreedhar Maskey, Yonas B. Dibike, Andreja Jonoski and Dimitri
Solomatine
What is the optimal pumping rate to
minimize time to cleanup?
P1 P2 P3
40 20 30
NEURAL NETWORK
G.O. TECHNIQUE
MINIMUM TIME: 3000 DAYS
TRAINING EXAMPLES FROM
MODFLOW/MODPATH
TRAINING EXAMPLES <<
SIMULATION ITERATIONS
< SECOND

33. Max Howarth - IBM Canada 2018
Groundwater Model Approximation with Artificial
Neural Network for Selecting Optimum Pumping
Strategy for Plume Removal
Shreedhar Maskey, Yonas B. Dibike, Andreja Jonoski and Dimitri
Solomatine
What is the optimal pumping rate to
minimize time to cleanup?
P1 P2 P3
40 20 30
NEURAL NETWORK
G.O. TECHNIQUE
MINIMUM TIME: 3000 DAYS
TRAINING EXAMPLES FROM
MODFLOW/MODPATH
TRAINING EXAMPLES <<
SIMULATION ITERATIONS
ORDER OF MAGNITUDE DECREASE IN TIME TO COMPLETE
< SECOND

34. WHAT WILL THE SALINITY LEVELS BE IN
14 DAYS?
RIVER MURRAY SYSTEM
ARTIFICIAL NEURAL NETWORKS IN HYDROLOGY. I: PRELIMINARY CONCEPTS
ARTIFICIAL NEURAL NETWORKS IN HYDROLOGY. II: HYDROLOGIC
APPLICATIONS
ASCE Task Committee on Application of Artificial Neural Networks in Hydrology

35. WHAT WILL THE SALINITY LEVELS BE IN
14 DAYS?
RIVER MURRAY SYSTEM

36. WHAT WILL THE SALINITY LEVELS BE IN
14 DAYS?
RIVER MURRAY SYSTEM
USE FLOW, WATER LEVEL,
DISCHARGE, AND TEMPERATURE
DATA FROM HERE
TO PREDICT SALINITY
HERE

37. WHAT WILL THE SALINITY LEVELS BE IN
14 DAYS?
RIVER MURRAY SYSTEM
USE FLOW, WATER LEVEL,
DISCHARGE, AND TEMPERATURE
DATA FROM HERE
TO PREDICT SALINITY
HERE

38. WHAT WILL THE SALINITY LEVELS BE IN
14 DAYS?
17 STATIONS
3 MEASUREMENT TYPES/STATION
~51 INPUT NODES
2 X 10 NEURON
HIDDEN LAYERS

39. WHAT WILL THE SALINITY LEVELS BE IN
14 DAYS?
1. GET THE DATA 2. PREPARE THE DATA
M1 M2 M3 … M50
0.8 -0.2 0.7 0.7 0.7TARGET FORMAT

40. WHAT WILL THE SALINITY LEVELS BE IN
14 DAYS?
1. GET THE DATA 2. PREPARE THE DATA
• Different stations collect different measurements
• Some measurements are blank…?
• Stations started collecting on different dates..!
• How far do I go back…??
• How do I transform the data…???
• Data needs to be normalized…!
M1 M2 M3 … M50
0.8 -0.2 0.7 0.7 0.7TARGET FORMAT

41. WHAT WILL THE SALINITY LEVELS BE IN
14 DAYS?
1. GET THE DATA 2. PREPARE THE DATA
• Different stations collect different measurements
• Some measurements are blank…?
• How far do I go back…??
• How do I transform the data…???
• Stations started collecting on different dates..!
• Data needs to be normalized…!
M1 M2 M3 … M50
0.8 -0.2 0.7 0.7 0.7TARGET FORMAT
Model Building
35%
Data Preparation
65%

42. Dataset and code are available for
you to try on box
ibm.biz/sabcs2018

43. Did it work?
10 EPOCHS

44. Did it work?
500 EPOCHS

45. Did it work?
1000 EPOCHS

46. Did it work? !
25000 EPOCHS

47. Did it work? !
25000 EPOCHS
SOME CONSIDERATIONS
• Network architecture was
largely arbitrary.
• Data cleansing was quick, and
not thorough.
• 1st iteration results - typically
do 10’s to 100’s
• Very promising capabilities.

48. Max Howarth - IBM Canada 2018

49. Max Howarth - IBM Canada 2018

50. Max Howarth - IBM Canada 2018

51. Max Howarth - IBM Canada 2018

52. Max Howarth - IBM Canada 2018

53. Max Howarth - IBM Canada 2018

54. Max Howarth - IBM Canada 2018

55. Max Howarth - IBM Canada 2018

56. A 3-dimensional
convolutional neural
network designed to work
with categorical and
continuous variables
Max Howarth - IBM Canada 2018
PREDICTIONS

57. © 2018 IBM Corporation IBM Services
THE RESULTS
REDACTED

58. Max Howarth - IBM Canada 2018

59. • Team made up of:
• Engineers (geological, materials, mining, etc.)
• Scientists (geophysicists, astrophysicists biologists,
etc.)
• Developers
• Data Scientists
• Like a startup backed by the power of IBM
• Always start with a Proof of Concept (PoC)
• Close working teams for successful
outcomes
10XRETURN ON INVESTMENT

60. 1. BE MINDFUL OF YOUR DATA
2. BUILD A USE CASE FOR THE DATA YOU HAVE
3. REPLACE LEGACY CONVENTIONAL MODELLING TECHNIQUES
4. START SMALL
5. DEFINE SUCCESS
6. GET A DATA SCIENTIST
Max Howarth - IBM Canada 2018

62. 62

63. © 2018 IBM Corporation IBM Services!63 IBM Services
WHY THIS USE CASE?
• Hand written documents have more variability.
• Hand written documents are typically lower
quality.
• Hand written documents do not have a consistent
orientation.
• Maps are coloured in.
• Maps are done underground, so documents are
often damaged.

64. © 2018 IBM Corporation IBM Services!64 IBM Services
USE CASE OBJECTIVES
1. Find a symbol
2. Classify a symbol
3. Find associated dip measurement
4. Read dip measurement
5. Determine strike angle
6. Georeference symbol
7. Read metadata

65. © 2018 IBM Corporation IBM Services!65
THE TECHNOLOGY: IMAGE CLASSIFIERS
IMAGE
CLASSIFIER

66. © 2018 IBM Corporation IBM Services!66
THE TECHNOLOGY: IMAGE CLASSIFIERS
IMAGE
CLASSIFIER
[cat, 0.98]

67. © 2018 IBM Corporation IBM Services!67
IMAGE CLASSIFIERS: TRAINING
IMAGE
CLASSIFIER
These are
all cats.

68. © 2018 IBM Corporation IBM Services!68
IMAGE CLASSIFIERS: TRAINING
IMAGE
CLASSIFIER
These are
all cats.
These are
all not
cats.

69. © 2018 IBM Corporation IBM Services!69
IMAGE CLASSIFIERS: TRAINING
IMAGE
CLASSIFIER
These are
all cats.
These are
all not
cats.
10K
of each

70. © 2018 IBM Corporation IBM Services!70
THE TECHNOLOGY: IMAGE CLASSIFIERS
IMAGE
CLASSIFIER
[cat, 0.98]

71. © 2018 IBM Corporation IBM Services!71
THE TECHNOLOGY: IMAGE CLASSIFIERS

72. © 2018 IBM Corporation IBM Services!72
THE TECHNOLOGY: IMAGE CLASSIFIERS
CONVOLUTION LAYER:
Learning complex patterns
from the input patterns.
POOLING LAYER:
Reduce spatial size of the
representation to reduce
amount of parameters and
computation in network

73. © 2018 IBM Corporation IBM Services!73
THE TECHNOLOGY: IMAGE CLASSIFIERS
CONVOLUTION LAYER:
Learning complex patterns
from the input patterns.
POOLING LAYER:
Reduce spatial size of the
representation to reduce
amount of parameters and
computation in network

74. © 2018 IBM Corporation IBM Services!74
THE TECHNOLOGY: IMAGE CLASSIFIERS
Convolutional Deep Belief Networks for Scalable Unsupervised
Learning of Hierarchical Representations. Lee, Ng
CONVOLUTION LAYER:
Learning complex patterns
from the input patterns.
POOLING LAYER:
Reduce spatial size of the
representation to reduce
amount of parameters and
computation in network

75. © 2018 IBM Corporation IBM Services!75
CHALLENGES
• Symbols are very small
• Everyone has different handwriting
• Orientation varies
• Colouring obscures symbol clarity
• Lots of unrelated symbols
• Huge variation in quality and condition of images

76. © 2018 IBM Corporation IBM Services!76
Image
Faster RCNN
(Main Object Detection)
Image Preprocessing
(Denoising, Enhancing
etc.)
CNN
(Detecting digits/alphabets
per contour)
Image Manipulation
(Cropping etc.)
Faster RCNN
(Digit Detection )
CNN
(Detecting angles for
each digit to get
orientation)
CNN
(Detect Orientation of
Objects)
Image Manipulation
(Cropping etc.)
Image Preprocessing
(Denoising, Enhancing
etc.)
Output
Information
(SymbolObjects,
geo-reference
coordinates,
metadata)
OCR
(Detecting digits/letters per
contour)
OCR
(Detecting digits/letters per
contour)
Image Manipulation
(Cropping etc.)
Symbol detection (Objective 1,2,3,4)
Coordinates detection (Objective 6)
Metadata Extraction (Objective 7)
Faster RCNN
OROBJECTIVES
1. Find a symbol
2. Classify a symbol
3. Find associated dip measurement
4. Read dip measurement
5. Determine strike angle
6. Georeference symbol
7. Read metadata

77. © 2018 IBM Corporation IBM Services!77
SOME MORE EXAMPLES

78. THE PROBLEM:
Target identification and
prioritization is expensive,
time consuming, and risky.

79. THE PROMISE:
Predictive models use large
volumes of historical data to
determine the likelihood of
future outcomes.

80. THE PROMISE:
Predictive models use large
volumes of geological data to
determine the likelihood of
mineralization.

81. © 2018 IBM Corporation IBM Services!81 IBM Services
LARGE VOLUMES
OF DATA
?
DATA QUALITY SUBJECTIVITY
WHY PREDICTIVE MODELLING?

82. © 2018 IBM Corporation IBM Services!82
MINERAL EXPLORATION CONDENSED
DATA DRIVEN
KNOWLEDGE DRIVEN
What do I know
about the
geological setting?
What does my
survey data tell
me?
• More data than a human can reasonably consume
• Can be affected by human bias
• Requires extensive experience & education
• Changes from person to person

83. © 2018 IBM Corporation IBM Services!83
MINERAL EXPLORATION CONDENSED
DATA DRIVEN
KNOWLEDGE DRIVEN
What do I know
about the
geological setting?
What does my
survey data tell
me?
• More data than a human can reasonably consume
• Can be affected by human bias
• Requires extensive experience & education
• Changes from person to person
SWEET SPOT
• Largely data driven models
• Tribal knowledge embedded in data representation
• Domain knowledge represented in model construction

84. © 2018 IBM Corporation IBM Services!84
MINERAL EXPLORATION PROCESS
SEARCH
• Data is usually disparate and silo’ed - geologists have
to aggregate it from multiple sources.
PREPARE
• Data is usually disparate and silo’ed - geologists have
to aggregate it from multiple sources.
MODEL
• Geologists examine the data, interpolate, and create
3D models to inform further exploration and mining
activities.

85. © 2018 IBM Corporation IBM Services!85
MINERAL EXPLORATION PROCESS
SEARCH
• Data is usually disparate and silo’ed - geologists have
to aggregate it from multiple sources.
PREPARE
• Data is usually disparate and silo’ed - geologists have
to aggregate it from multiple sources.
MODEL
• Geologists examine the data, interpolate, and create
3D models to inform further exploration and mining
activities.
70%
OF A GEOLOGIST’S TIME
UP TO

86. © 2018 IBM Corporation IBM Services
PREDICTIVE MODELLING FRAMEWORK
PREREQUISITES FOR AI
• Large quantity of data
• Data is cleaned
• Data is structured and organized
• Business objectives are understood

87. © 2018 IBM Corporation IBM Services
PREDICTIVE MODELLING FRAMEWORK
PREREQUISITES FOR AI
• Large quantity of data
• Data is cleaned
• Data is structured and organized
• Business objectives are understood
70%
OF A DATA SCIENTIST’S TIME

88. © 2018 IBM Corporation IBM Services
PREDICTIVE MODELLING FRAMEWORK

89. © 2018 IBM Corporation IBM Services
PREDICTIVE MODELLING FRAMEWORK

90. © 2018 IBM Corporation IBM Services
PREDICTIVE MODELLING FRAMEWORK
AI HELPS GEOLOGISTS
WORK BETTER HERE
SO MORE TIME CAN BE
SPENT HERE

91. © 2018 IBM Corporation IBM Services
PREDICTIVE MODELLING FRAMEWORK
AI INFORMS THE MODELLING PROCESS BY
ALLOWING EXPERIMENTS TO BE TESTED

92. THE PROMISE:
Predictive models use large
volumes of geological data to
determine the likelihood of
mineralization.

93. © 2018 IBM Corporation IBM Services!93
SPEND LESS TIME PREPARING DATA
3D GIS PLATFORM FOR
MODELLING
• Data is aggregated, correlated, and stored
in a manner that is conducive to both
geological and predictive modelling.
NON-INVASIVE
• Continue to use your existing tools and
software to collect data - no breaking of
business processes.
NECESSARY
• Clean, organized data is a requirement for
modelling - maximize the value of your
preparation activities.

94. © 2018 IBM Corporation IBM Services!94
SPEND LESS TIME PREPARING DATA
3D GIS PLATFORM FOR
MODELLING
• Data is aggregated, correlated, and stored
in a manner that is conducive to both
geological and predictive modelling.
NON-INVASIVE
• Continue to use your existing tools and
software to collect data - no breaking of
business processes.
NECESSARY
• Clean, organized data is a requirement for
modelling - maximize the value of your
preparation activities.

95. THE PROMISE:
Predictive models use large
volumes of geological data to
determine the likelihood of
mineralization.

96. © 2018 IBM Corporation IBM Services!96
PREDICTIVE MODELLING REQUIREMENTS
Criteria
1. Consume large amounts of data.
2. Can use geospatial information (i.e. work in 3D).
3. Can use categorical variables.
4. Low requirement for knowledge engineering.
5. Can be trained on a specific area (e.g. brownfield).
6. Can be trained on a non-specific area (e.g. greenfield)

97. © 2018 IBM Corporation IBM Services!97
CONVOLUTIONAL NEURAL NETWORKS
Criteria
1. Consume large amounts of data.
2. Can use geospatial information (i.e. work in 3D).
3. Can use categorical variables.
4. Low requirement for knowledge engineering.
5. Can be trained on a specific area (e.g. brownfield).
6. Can be trained on a non-specific area (e.g. greenfield)

98. © 2018 IBM Corporation IBM Services!98
WATSON FOR GEOLOGY PREDICTIVE MODELS
A 3-dimensional convolutional neural
network designed to work with
categorical and continuous variables
Network architecture designed with
recognition of high level geological
features in mind - training the model
adds further context.

99. © 2018 IBM Corporation IBM Services!99
FEATURE ENGINEERING & DATA EXTRACTION
Can we teach the model about ternary
diagrams?
Can we leverage even more data?
DRILL LOG COMMENT COMPREHENSION
MAP ANALYSIS

100. © 2018 IBM Corporation IBM Services!100
NEXT STEPS
GREENFIELDS BROWNFIELDS RESOURCES EXPANSION
MODEL TESTEDQ1 2019