Contrasting faulting styles of salt domes and volcanoes: Can unsupervised learning techniques differentiate fault styles?

Abstract

An attractive feature seen on seismic data, also known as funny-looking thing (FLT), has a wide range of interpretations, from noise patterns to amplitude anomalies. An example of an FLT is the similar faulting patterns between a volcanic intrusion and a salt intrusion from the point of view of a machine learning (ML) algorithm. Oftentimes, seismic interpreters do not have a complete data set or geologic background to determine the genesis of the observed features. This can be particularly perplexing when trying to determine if an intrusion is volcanic or halokinetic in origin because they exhibit similar geomorphologies. Examining the differences in these features in the Gulf of Mexico, a well-documented salt basin, and the Taranaki Basin in New Zealand, which is igneous prone, can help to understand the differences. The analysis aims to discern geologic features based on the geometries and attributes shared by seismic data and remote sensing tools. Seismic attributes and ML techniques highlight differences and similarities between the intrusions, including the discussion of using ML techniques, such as self-organized maps (SOMs), an unsupervised ML technique, and cluster fault systems, without regard to the geologic context. The attributes used in the SOM are fault probability, fault dip azimuth, fault dip magnitude, and thin-bed detector. Fault probability is performed through a combination of convolutional neural network fault prediction and a skeletonization process. Once the faults are clustered using SOM, the visualization of fault architecture due to the existing mount (either volcano or salt dome) is done considering high fault probabilities (>75%). The methodology consists of selecting the neurons from the SOM grid corresponding to the presence of faults and combining them with fault probability and a fault dip azimuth using a crossplot. The crossplot product assists in the automatic extraction of the fault planes using: (1) a voxel representation of the fault planes and (2) fault patches representing the fault planes. Moreover, the visualization technique defined demonstrates that the crossplot product yields better-defined fault planes. With the fault system characterized, compared horizon slices using coherence, fault dip magnitude, and azimuth against remote sensing images with similar attributes. In conclusion, our methodology combines technologies to differentiate the genesis of intrusion — salt or igneous — using the fault presence and could be helpful in frontier exploration or planetary exploration. Geologic feature: Comparison of extensional faults due to salt and igneous intrusions Seismic appearance: In the symmetric salt dome, crown faulting style with a set of synthetic and antithetic faults around the salt dome; for the igneous systems, symmetric radial faults Formation: Upper and Lower Wilcox (Gulf of Mexico); Ariki Formation — Mo hakatino Formation (Taranaki) Age: Upper Miocene and Lower Pliocene (Gulf of Mexico); Mid Pliocene (Taranaki) Location: Offshore Gulf of Mexico and Offshore Northeast New Zealand Seismic data: From the Gulf of Mexico: B-78-89-LA-3D, B-69-94-LA-3D, B-54d-94-LA-3D, and B-69a-94-LA-3D. From New Zealand: Nimitz and Kora 3D Analysis tools: Structural-oriented filter, Sobel similarity, fault dip magnitude and azimuth, CNN automatic fault prediction, and SOM classification