Seismic interpretation involves more than simply picking faults and horizons. It involves the interpretation of geologic features — their geometry, morphology, and the context of one group of rocks to another. It involves using well log information, memories from fieldwork, and photos from outcrops. It involves the understanding of salt mechanics, wave propagation, and signal analysis. It requires context and agile minds that can readily distinguish mud volcanoes from salt diapirs or multiples from reflectors. It is a difficult practice, and individuals spend their entire careers devoted to it.

Seismic attributes have always been considered by many to be an art form — a “dark art” — practiced by a chosen few. The proliferation of attributes to the workstation has not, unfortunately, proliferated the understanding of what the attributes mean or of what they are capable. Today, you will often find the seismic attribute specialists in quantitative interpretation or computational geophysics groups. The perspective of these specialists and of general interpreters can be quite different. They understand both physics and geology in different ways and at different levels.

As the discipline moves toward new technologies and the promises of new algorithms like convolutional neural networks and other forms of machine learning, we must remind ourselves that the technical understanding required of scientists and professionals grows accordingly. However, like the proliferation of seismic attributes (e.g., geometric and single-trace), machine learning approaches will feel underwhelming by those who fail to understand both the algorithms and what can reasonably be achieved.

This dissertation provides the reader with the foundational knowledge one requires to begin to understand seismic attributes and how they can be used with machine learning algorithms. I begin by establishing a common framework on which to communicate. I build upon that through the development of a procedure to enhance faults in seismic data using commercially available tools, and I end with the introduction of a simple, but effective, use of self-organizing maps, a simple machine learning algorithm.