This document summarizes the application of machine learning in the oil and gas industry. It discusses how the author participated in an SEG machine learning challenge to classify lithofacies from well log measurements. Several algorithms were tested including SVM, logistic regression, random forest and gradient boosting. The top team used feature engineering, one vs one multiclass modeling with XGradientBoost and model tuning. Future work includes improving the classifier with more geological data from experts and combining algorithm and expert rankings.

1. Application of Machine
Learning in Oil and Gas
Industry
Priyanka Raghavan

2. About me
 Team Lead at Schlumberger working on Prestack Seismic Interpretation product
 Did Coursera Machine Learning Course
 Little experience on Machine Learning before this challenge
 Participated in SEG “machine learning challenge” with another team member (Steve
Hall)

3. Agenda
 Background oil & gas Industry
 Problem statement
 Exploring Data
 Algorithms we tried
 What worked?
 What did the top team do?
 Future Work and Lessons Learnt

4. Background- Oil & Gas (Big Data)
 Big data in Oil and Gas Industry
 Well logging was discovered by Marcel and Conrad Schlumberger in 1927
 Nearly 100 years of oil field data
 Typical data size- Normal to see peta-byte data
 Many tools and techniques in Geoscientist toolbox for identifying Lithofacies
 Aim to find reservoir rock or target areas for hydrocarbon extraction
 What is Well logging?
 What is Lithofacies and classification?

5. Bring Machine Learning to Geoscientist
toolbox
 Society of Exploration Geophysict(SEG)
conducted a challenge
 Identify a Lithofacies based on well log
measurement
 Team came 5th in competition
 Used open source tools like python, jupyter
notebook, scikit-learn library and Git for version
control
 Exposed to various algorithms through the
challenge

6. Dataset
 Dataset is in Kansas region
 3232 training data points
 Well log measurements
 Gamma Ray intensity-> Measures radioactivity of surrounding rocks
 Resistivity- >Saturation of oil,gas,water
 PhotoElectric-> Measures photo electric absorption
 Neutron-density porosity difference-> Porosity indicator
 Average neutron-density porosity -> Porosity indicator
 Nonmarine/marine indicator
 Relative position
 Facies data available from 9 wells
 Goal was to classify blind data with only measurements

7. Algorithms tried for facies classification
 SVM
 Logistical Regression
 K nearest neighbours
 Random forest
 Majority Voting
 Gradient boosting
 One vs One Multiclass GBM- Winning solution and final one

8. What did we do that worked?
 Feature engineering
 One vs One multiclass using Random Forest
 No abrupt boundaries between facies
 Intelligent replacement of PE values, which had nulls
 Kaggle type approach helped as we learnt from other teams

9. What the Winning Team did & Lessons
Learnt
 Better Feature extraction
 Finding new features
 Model Tuning
 Using XGradientBoost
 Some key takeaways from our experience (to be elaborated)
 Machine Learning is a good strategy for facies prediction

10. Future work
 Get some more geological information by collaborating with experts
 Improving classifier by training on other datasets
 Concentrate on weighting by being more intelligent
 Rank results from algorithm and Geoscientist, like a Turing test?