Seismic Facies Classification using Generative Topographic Mapping – A Case Study from Offshore Nova Scotia

Abstract

Summary One of the major objectives of seismic interpretation is to effectively predict the distribution of reservoir facies away from well control. With the advent of increasing number of meaningful seismic attributes, it is time consuming and laborious to analyze them through conventional analytical methods. Machine learning techniques analyze higher dimensional data points faster and effectively. Automated seismic facies classification techniques are increasingly becoming important in identifying the potential hydrocarbon bearing zone and favorable facies. Such facies classification techniques, or, automated clustering algorithms, help arrange similar seismic traces based on the waveform shape, amplitude, phase, frequency, and other relevant seismic attributes. The main objective of automated facies classification using machine learning techniques is to perform facies classification fast and efficiently using several relevant seismic attributes for mapping the facies distribution and effective identification of the sweet spots. The automated clustering algorithms fall into two categories – supervised and unsupervised algorithms. Unsupervised machine learning algorithms are purely data driven and help in recognizing and classifying the patterns from a dataset without any a priori information. A posteriori information such as well data, is integrated into the results for recognizing the facies classification and calibrating the interpretation. Unsupervised learning methods also help to highlight subtle stratigraphic features that might otherwise be unnoticed using conventional analytical methods. In this study, we adopted a recent unsupervised classification technique called, generative topographic mapping (GTM). We applied this technique to a dataset from Offshore Nova Scotia, to extract the natural clusters from the seismic data for facies classification. Using the GTM technique applied to seismic data, we were able to map the distribution of different facies and potential sweet spots in the study area.