From clusters to facies: Integrating domain knowledge into ML-based seismic interpretation in deepwater channels

Abstract

Machine learning workflows in geoscience are increasingly capable of identifying statistical patterns in complex, multiattribute datasets, but pattern recognition alone does not constitute geologic interpretation. Domain knowledge must be deliberately built into every stage of the workflow, from attribute selection through model evaluation, not applied as an afterthought to algorithm outputs. An integrated ML-geologic framework is presented here, demonstrating how unsupervised cluster model outputs are evaluated and interpreted through a two-tier framework that balances statistical performance with geologic interpretability, combining multiattribute seismic analysis, kernel principal component analysis, and unsupervised clustering. Applied to the Romney 3D seismic volume, Taranaki Basin, New Zealand, the workflow identifies nine seismic facies: three response-type facies calibrated at Romney-1 and six stratigraphic interval facies defined in cross-section. Architectural element mapping reveals a multi-phase deepwater channel system with evidence of lateral migration, channel abandonment, and aggradational stacking, with cluster distributions validated across timeslice, chair view, and cross-sectional perspectives. The workflow treats machine learning as a pattern recognition tool whose outputs must be evaluated against geologic reality, unsupervised in algorithm, geologically supervised in application, and is transferable to any seismic facies identification problem where multiattribute analysis is warranted and well control is limited, provided attribute selection rationale and domain knowledge are rebuilt for each application.