Seismic characterization with unsupervised machine learning applications for facies classification

Abstract Machine learning (ML) techniques hold significant promise in identifying and delineating heterogeneous 3D seismic facies because they allow us to integrate the information content from multiple seismic attribute volumes. ML methods can be broadly categorized into four types: unsupervised learning, which uses the seismic attributes themselves as the training data and the data to be analyzed; supervised learning, which relies on labeled data — classified using well control, production data, or expert interpretation — for training; hybrid learning, which combines aspects of supervised and unsupervised approaches; and deep learning, which uses neural networks trained on large, labeled datasets to extract complex patterns from seismic data. However, great care must be taken not only in selecting but also in properly scaling the attributes to be used. Sometimes, these exercises are carried out mechanically, leading to compromised interpretations and discouraging results. In this study, we examined two well-established unsupervised ML techniques: principal component analysis and self-organizing mapping, applied to a seismic data volume from New Zealand. We found that these ML methods can enhance spatial resolution and, therefore, assist in seismic interpretation. However, it is crucial for the interpreter to ensure that the appropriate ML method is chosen, the assumptions underlying the applied techniques are met, and any available options, such as weighting of input attributes, are considered and used effectively.