Round Corner
Department of Computer and Information Science


Automatic Fault Surfaces Extraction from Seismic Volume using Machine Learning

This project is associated with computational sciences activity group under the collaboration framework between Schlumberger, the largest oilfield services company in the world, and IPT (Petroleum Engineering & Applied Geophysics) and IDI (Computer and Information Science) at NTNU. The aim is to automatically extract fault surfaces from seismic volume with human intervention as less as possible.
Interpreting faults in 3D seismic data is still one of the most time-consuming and tedious aspects of seismic interpretation. Automatic Fault Extraction (AFE) is a process designed to automatically interpret fault surfaces from 3D seismic volume. The AFE workflow is comprised of the following three main steps explicitly or implicitly:
A. Fault attributes: select an appropriate fault-sensitive attribute to highlight the fault location
B. Fault likelihood: transform the attribute volume into a fault likelihood/confidence volume
C. Fault surfaces: generate a localized surface mainly in form of polygons, from the confidence volume
Over the last decade, a multitude of AFEs have been developed, including kinds of fault attributes (e.g. chaos, curvature), various fault likelihood (ant tracking, hough transform, skeletonization, applying geological constrain) and different fault surfaces (e.g. in format of fault sticks, polygon meshes). Although a few of them has been implemented in the commercial software but it is not satisfying the end users.
Exploring new methods to further improve the performance and reduce the human intervention accelerates the interpretation workflow by orders of magnitude. Thus it continuously attracts lots of researchers from kinds of disciplines and try to treat the AFE from different viewpoints. With our background in computer science, it would be possible to at least investigate machine learning for fault extraction. A choice from machine learning would be deep neural network (DNN). It uses a cascade of many layers of nonlinear processing units for feature extraction and transformation. Each successive layer uses the output from the previous layer as input. The algorithms may be supervised or unsupervised and applications include pattern analysis (unsupervised) and classification (supervised). In the context of fault extraction, the classification is fault or not fault.
The project is ambitious and could lead to a publication. It will be implemented as a Plugin for Petrel, which is a Schlumberger owned E&P software platform that provides an integrated solution from exploration to production.
[Necessary] Familiar with Linux, Good programming skills in Python, TDT4195 (Visual Computing Fundamentals)
[Preferably] TDT4230 (Graphics & Visualization), TDT4265 (Computer Vision), Machine Learning.
Dataset to be used: Will be provided by Schlumberger
Advisors: Liyuan Xing (NTNU), Victor Aarre (Schlumberger), Theo Theoharis (NTNU).
If interested, please contact Liyuan Xing,



Theoharis Theoharis Theoharis Theoharis
422 IT-bygget
735 91447 
NTNU logo