In Core Classification

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 28

Application of

Machine Learning to
Facies Classification
of Carbonate Core
Images
Sharinia Kanagandran1, Dr Cedric John

1sharinia@hotmail.com

1
Why – Why is Digitalisation important?
1. Increasing Global Energy Demand

3D
Seismic

(IEA, 2011) (Norwegian Petroleum Directorate, 2018)

2
Why – Why is applying Machine Learning (ML) to core significant?
2. Proof of Concept
Seismic Core? Wireline

IBM Schlumberger Research since


120 days to 3 days 1980s
80% seismic interpretation time
seismic interpretation time

(IBM, 2018; Schlumberger, 2018) 3


Why – Why apply ML to facies* classification?
3. Classification Inconsistency 2 out of 3 *facies is used as a short term for Dunham texture
geologists would disagree with
each other on classification of
carbonate thin sections

Wackestone Thin Section

2mm

(Lokier & Junaibi, 2015) 4


Aims of the Study

1. Evaluate Random Forest (RF) and Convolutional Neural Networks (CNNs) for
facies identification in high resolution carbonate core images using Dunham
classification.

2. Focusing on CNNs, assess the various architectures and methods to find the
most suitable CNN configuration for facies classification.

3. Compare classification by machine learning algorithms to classification by


geologists.

5
Outline
v Why ü
vAims of the Study ü
v Machine Learning Primer ß
vDataset
v Methodology – Workflow
- Random Forest
v Random Forest Method & Results
v Methodology – Convolutional Neural Networks (CNNs)
v CNN Results
v Discussion – Potential Implementation in the Petroleum Industry
- Investigating Accuracy and Error in CNNs
- Can Geologists surpass Machine Learning Algorithms?
- Can the Algorithm Pass for a Geologist
v Conclusions 6
Machine Learning Primer
“Field of study that gives computers the ability to learn without being explicitly programmed”
(Arthur Samuel, 1959)

(Mathworks, 2018) 7
Dataset

(John & Mutti, 2005) 8


Dataset
Site 1196

Age: Miocene to Pleistocene


Environment of deposition: Platform to slope carbonates
Facies Classified: Mudstone to Rudstone (7 facies)
Complexities: - Heavily dolomitised intervals in Site 1196A
- Colour variations in dolomite
- Core Fragmentation (Shipboard Scientific Party, 2011) 9
Methodology - Workflow
Input Data

1.5m
1. Preprocess Data 2. Classify All Data

Floatstone

3. Train & Test Algorithm


(RF/CNN)

Modify Hyperparameters
4.5cm
based on Accuracy Maximum Accuracy
Model
4. Output Accuracy of
Algorithm & Confusion Applications of Model:
Low High
Matrix Accuracy - Testing on Maldives data
Accuracy
- Classification of Unlabelled
Core

10
Methodology – Random Forest
A supervised learning algorithm which builds a forest from an ensemble of Decision Trees. Input Data:
1000 pixels

1000
= 10$ %&'()*
pixels

4.5cm

(Kumar, 2016) 11
Random Forest Methods – RF implementation on Scikit-Learn

12
Random Forest Methods – RF implementation on Scikit-Learn
No. of images No. of Images
per facies: per facies:

Facies:
Crystalline Dolomite
Grainstone
Floatstone
Rudstone

4 4 4 4
No. of Facies 13
Methodology – Convolutional Neural Networks (CNNs)

14
Potential Implementation in the Petroleum Industry

60x
faster

10x
higher
resolution

15
CNN Results
CNN Results (Core) CNN Architecture

No. of Images
per Facies:

Highest accuracy
model for 7 facies
classification
CD, F, CD, F, CD, F, CD, F, CD, F, CD, F, CD, F, CD, F,
CD, F, CD, F, G, M, CD, R G, M,
G, M, G, M, G, M, G, M, G, M, G, M,
G, R G, M, R
P, R, W P, R, W P, R, W P, R, W P, R, W P, R, W P, R, W P, R, W

No. of Facies 16
Investigating Accuracy and Error in CNN

17
Can Geologists surpass Machine Learning Algorithms?

18
Classification Inconsistency among Geologists

Floatstone

19
(Lokier & Junaibi, 2015)
Can the Algorithm Pass for a Geologist?
“Select which facies you think the human assigned”

20
Maldives Data

Marion Plateau Maldives

Rhodolith

Fossil Diversity:
Rhodoliths Fossil Diversity:
Benthic foraminifera Encrusting Algae
Coralline Algae Branching Corals
Gastropods
Bivalves

21
Conclusions
1.Convolutional Neural Network is a superior algorithm to Random
Forest for facies classification of carbonate core images and achieves
89.2% accuracy for 7 facies classification.

2.Experienced geologists are slightly more accurate but are


inconsistent, whereas the algorithm is consistent, 60 times faster and
∼10 times higher resolution.

3.Both the algorithm and geologists display the human cognitive bias,
saliency. Moreover, the algorithm passes for a geologist.

4.We recommend the petroleum industry adopt machine learning for


carbonate core classification.

22
Future Works
• Train algorithm on a more globally heterogeneous dataset

• A survey should be conducted to test the accuracy of geologists with


physical core classification relative to the algorithm’s classification of
high-resolution images

• Additional Studies; Image Segmentation with U-Net, Fossil


Identification with the Algorithm, Mobile app using MobileNet for
drillers/engineers

23
Future Works

Image Segmentation Fossil Identification


• Aim – Quantitative method which would generate
objective and unbiased classification.
• Removes the need for multiple experienced
geologists to classify each sample
• Eliminates human performance and human bias
• Quantitative data could be correlated with
porosity and permeability

24
Future Works

Image Segmentation Fossil Identification


Rhodolith • Aim – Identification of fossils – particulary key
index fossils for biostratigraphy markers.
• Once fossils are segmented, a second algorithm
can identify the fossils
• Creates a thorough algorithm which extracts the
most important information from core data
• Can be used for well correlation
• Identification of fossil proportions

25
Acknowledgements

I’d like to thank Cedric John for his


incredible guidance and
encouragement throughout this
project. His feedback and ideas
have been invaluable to this thesis.

“Technology is not going to replace


geoscientists, it lets us focus on the less
obvious and more impactful
interpretation”
- Lindsey Lomas, Schlumberger SIS (2018) 26
References
IBM, 2018, AAPG Energy Transition Conference.

International Energy Agency, 2011, World Energy Outlook 2011.

John, C. and Mutti, M. (2005). Relative Control of Paleoceanography, Climate, and Eustasy over Heterozoan Carbonates: A
Perspective from Slope Sediments of the Marion Plateau (ODP LEG 194). Journal of Sedimentary Research, 75(2), pp.216-230.

Kumar, N., 2016, Random Forest Algorithm, An Interactive Discussion: <https://www.linkedin.com/pulse/random-forest-algorithm-


interactive-discussion-niraj- kumar/> (accessed August 23, 2018).

Lokier, S. W., and M. Al Junaibi, 2016, The petrographic description of carbonate facies: are we all speaking the same language?
Sedimentology, v. 63, no. 7, p. 1843–1885, doi:10.1111/sed.12293.

MathWorks, 2018, What Is Machine Learning? | How It Works, Techniques & Applications - MATLAB & Simulink:
<https://www.mathworks.com/discovery/machine-learning.html> (accessed September 9, 2018).

Norwegian Petroleum Directorate, 2018, Resource Report - Exploration 2018: 7 p.

Schlumberger, 2018, AAPG Energy Transition Conference.

Shipboard Scientific Party, 2001, Ocean Drilling Program Leg 194 Preliminary Report Sea Level Magnitudes Recorded by Continental
Margin Sequences on the Marion Plateau, Northeast Australia.
27
Conclusions
1.Convolutional Neural Network is a superior algorithm to Random
Forest for facies classification of carbonate core images and achieves
89.2% accuracy for 7 facies classification.

2.Experienced geologists are slightly more accurate but are


inconsistent, whereas the algorithm is consistent, 60 times faster and
∼10 times higher resolution.

3.Both the algorithm and geologists display the human cognitive bias,
saliency. Moreover, the algorithm passes for a geologist.

4.We recommend the petroleum industry adopt machine learning for


carbonate core classification.

28

You might also like