|
" Seismic and Well Log Based Machine Learning Facies Classification in the Panoma-Hugoton Field, Kansas and Raudhatain Field, North Kuwait "
Dwihusna, Nadima
Jin, Ge
| Document Type
|
:
|
Latin Dissertation
|
| Language of Document
|
:
|
English
|
| Record Number
|
:
|
1053956
|
| Doc. No
|
:
|
TL53073
|
| Main Entry
|
:
|
Dwihusna, Nadima
|
| Title & Author
|
:
|
Seismic and Well Log Based Machine Learning Facies Classification in the Panoma-Hugoton Field, Kansas and Raudhatain Field, North Kuwait\ Dwihusna, NadimaJin, Ge
|
| College
|
:
|
Colorado School of Mines
|
| Date
|
:
|
2020
|
| Degree
|
:
|
M.S.
|
| student score
|
:
|
2020
|
| Note
|
:
|
164 p.
|
| Abstract
|
:
|
This thesis focuses on applying machine learning on facies classification presented in three case studies: 1) supervised, 2) semi-supervised, and 3) unsupervised machine learning to classify facies in various well logs and seismic data in the Hugoton-Panoma Field, Kansas and Raudhatain Field, North Kuwait. The first study applies supervised machine learning to a set of labeled well logs from the Hugoton-Panoma Field, Kansas. The Hugoton-Panoma field is a stratigraphic trap overlying monocline, with a primary gas reservoir rock found in the Permian dolomite in the Chase and Council Grove group. The supervised machine learning algorithms consist of 2D Convolutional Neural Network (CNN), K-Nearest Neighbors (KNN), Support Vector Machine (SVM), Random Forest, and Multilayer Perceptron (MLP). The algorithms classified the reservoir lithofacies sequences and updated the geologic interpretation of the dolomite target reservoir intervals. The supervised algorithms perform best with optimized hypoparameters and balanced training set. Supervised machine learning methods also tend to perform more accurately with more differential facies to classify. Once these supervised machine learning algorithms are fully optimized and trained, facies classification using machine learning is approaching the accuracy of traditional interpretation methods. The second and third case studies use unlabeled well log and post-stack seismic data to classify and characterize the facies variations in the deep (14,000 ft) Jurassic reservoirs of the Raudhatain Field, North Kuwait. The heterogeneous facies and strong seismic interbeded multiples affects the reservoir section. The reservoir interval is immediately below thick layers of Hith-Gotnia Formations with alternating salts and anhydrite, high-pressure high-temperature and sour fluid conditions provide geomechanical and environmental challenges. The Najmah Kerogen and Marrat formations are the main resource play development project in Kuwait, and the reservoir characterization is still uncertain. Hence, integrating machine learning to perform facies classification is essential to build a better understanding of subsurface conditions for further field development planning and exploration. The second case study involves semi-supervised learning for facies classification to the unlabeled well log data in the Raudhatain Field, North Kuwait. The K-Means unsupervised learning algorithm was trained and petrophysics domain knowledge was applied to label the classes. Combining petrophysics-based domain knowledge with machine learning allows the algorithm to be scalable to larger datasets, increase efficiency, and assist interpretation. Through this study, the reservoir characterization has been improved in the Upper Jurassic. Semi-supervised machine learning algorithm has classified the Hith-Gotnia interval of salt and anhydrites facies variation, and the Najmah to Marrat formations which contains the kerogen, siltstones, limestones, carbonates, and sandstone depositions. The third case study involves unsupervised machine learning for facies classification to the unlabeled post-stack seismic volume in the Raudhatain Field, North Kuwait. First, instantaneous and geometric attributes were generated from the post-stack seismic data. Through the Principal Component Analysis (PCA), a suitable combination of the attributes are selected. Afterwards, the Self-organizing Map (SOM) analysis was applied to identify the neurons or neural clusters visualized in a 2D Color Map. Each neuron in the 2D Color Map represents a cluster of data points. SOM is a good seismic visualization tool for interpreters to reveal additional information in the seismic data that may lead to geologic findings. In this study, facies recognition through SOM performed successfully in the Upper Jurassic, and limited success for the Lower Jurassic due to the complexity of the overburden and quality of the seismic data. In summary, machine learning improves the efficiency and aids the task of interpreters for exploring and characterizing facies in large volumes of well log and seismic data. All three case studies provides valuable insights in applying different types of machine learning for geologic interpretation of each fields. Facies recognition enabled by machine learning has high potential in the future of reservoir characterization.
|
| Descriptor
|
:
|
Artificial intelligence
|
|
|
:
|
Computer science
|
|
|
:
|
Geophysics
|
| Added Entry
|
:
|
Jin, Ge
|
| Added Entry
|
:
|
Colorado School of Mines
|
| |