Hypothesis-based machine learning for deep-water channel systems
Date
2020
Authors
Vento, Noah Francis Ryoichi, author
Stright, Lisa, advisor
Ronayne, Michael, committee member
Anderson, Charles, committee member
Journal Title
Journal ISSN
Volume Title
Abstract
Machine learning algorithms are readily being incorporated into petroleum industry workflows for use in well-log correlation, prediction of rock properties, and seismic data interpretation. However, there is a clear disconnect between sedimentology and data analytics in these workflows because sedimentologic data is largely qualitative and descriptive. Sedimentology defines stratigraphic architecture and heterogeneity, which can greatly impact reservoir quality and connectivity and thus hydrocarbon recovery. Deep-water channel systems are an example where predicting reservoir architecture is critical to mitigating risk in hydrocarbon exploration. Deep-water reservoirs are characterized by spatial and temporal variations in channel body stacking patterns, which are difficult to predict with the paucity of borehole data and low quality seismic available in these remote locations. These stacking patterns have been shown to be a key variable that controls reservoir connectivity. In this study, the gap between sedimentology and data analytics is bridged using machine learning algorithms to predict stratigraphic architecture and heterogeneity in a deep-water slope channel system. The algorithms classify variables that capture channel stacking patterns (i.e., channel positions: axis, off-axis, and margin) from a database of outcrop statistics sourced from 68 stratigraphic measured sections from outcrops of the Upper Cretaceous Tres Pasos Formation at Laguna Figueroa in the Magallanes Basin, Chile. An initial hypothesis that channel position could be predicted from 1D descriptive sedimentologic data was tested with a series of machine learning algorithms and classification schemes. The results confirmed this hypothesis as complex algorithms (i.e., random forest, XGBoost, and neural networks) achieved accuracies above 80% while less complex algorithms (i.e., decision trees) achieved lower accuracies between 60%-70%. However, certain classes were difficult for the machine learning algorithms to classify, such as the transitional off-axis class. Additionally, an interpretive classification scheme performed better (by around 10%-20% in some cases) than a geometric scheme that was devised to remove interpretation bias. However, outcrop observations reveal that the interpretive classification scheme may be an over-simplified approach and that more heterogeneity likely exists in each class as revealed by the geometric scheme. A refined hypothesis was developed that a hierarchical machine learning approach could lend deeper insight into the heterogeneity within sedimentologic classes that are difficult for an interpreter to discern by observation alone. This hierarchical analysis revealed distinct sub-classes in the margin channel position that highlight variations in margin depositional style. The conceptual impact of these varying margin styles on fluid flow and connectivity is shown.