Deepwater seismic facies and architectural element interpretation aided with unsupervised machine learning techniques: Taranaki basin, New Zealand

The use of unsupervised machine learning (ML) methods such as self-organizing maps (SOMs), has gained a significant foothold within the seismic interpretation community to enhance results and help identify similar patterns in the data. Workflows for each geological setting become, however, necessary. We analyzed a series of geometrical, instantaneous, spectral, and textural seismic attributes to provide multi-attribute input combina-tions that, with SOMs, allow for a proper interpretation of architectural elements in deepwater settings. By studying the Miocene section within the Pipeline 3D seismic dataset in southern Taranaki Basin, we show that GLCM entropy, RMS amplitude, sobel filter similarity, and frequencies seismic attributes used as inputs into the SOM enhance the interpretation of the features of interest, initially overlooked by single attributes analysis. This method had not been applied in the area yet. Results allow understanding and delineating architectural elements of deepwater systems, such as mud-and sand-filled channels, mass transport deposits (MTDs), and marine shales in the basin floor. Additionally, more minor features such as levees, overbank deposits, and sand waves are better mapped. The use of the Pukeko well helped to understand the system's evolution, which we interpret as a fan system incised by younger, more sinuous channels. Overall, the system turns more mud-prone as it becomes younger (upwards). Our study contributes to understanding the paleogeomorphology of the area, which can help optimize drilling plans, whereas our method offers three input attribute combinations for SOMs that can be applied in similar geological settings to aid in the interpretation optimization process.