UNSUPERVISED SEISMIC FACIES CLASSIFICATION IN A DEEPWATER CHANNEL SYSTEM: DIMENSIONALITY REDUCTION, CLUSTER VALIDATION, AND STRATIGRAPHIC FRAMEWORK DEVELOPMENT FOR MACHINE LEARNING INTERPRETABILITY.

Abstract

Machine learning (ML) techniques are increasingly adopted in seismic interpretation workflows, yet their rapid integration has outpaced the development of frameworks for evaluating whether their outputs are geologically meaningful. Too often, ML outputs are treated as geologic answers rather than as one step in a broader interpretation process, optimizing for statistical performance rather than geologic validity. This dissertation develops and demonstrates a practical framework for integrating unsupervised ML into seismic facies interpretation within a geologically informed context, using a deepwater channel system in the Taranaki Basin, New Zealand as the primary study area. Key findings demonstrate that dimensionality reduction (DR) improves both visual interpretability and unsupervised clustering performance across 55 model configurations — but that statistical performance alone is insufficient for model selection. The best-performing statistical model is not necessarily the most geologically meaningful one. A robust geologic foundation, including systematic seismic visualization, understanding of depositional processes, and stratigraphic context, must precede and inform ML evaluation. The central contribution of this work is a transferable framework that guides practitioners through attribute selection, DR technique selection, model evaluation, and geologic validation. The framework offers a disciplined counterbalance to the pressure of rapid ML adoption: the algorithm answers what you ask, not what you mean. The goal is not to replace geologic reasoning with computation, but to demonstrate how ML, when implemented thoughtfully within a geologically informed workflow, can enhance interpretive resolution and increase confidence in geologic models.