Kelly Bradbury and Tony Lowry

Document Type


Publication Date


First Page


Last Page



Borehole geophysical data provide important in situ observations for identifying and characterizing geologic and structural features in the subsurface. We perform an unsupervised classification of geophysical logs from the Sand Andreas Fault Observatory at Depth (SAFOD) borehole in order to define intervals of distinct geophysical properties. We focus on borehole geophysical data collected during Phase 3 drilling operations from 3.0-3.3 km measured depth, which encompasses the active trace of the San Andreas Fault. We use dimensionality reduction to increase the interpretability of the retrieved clusters, and we compare the performance of distance- and density-based clustering techniques. K-means clustering produces highly fragmented and overlapping clusters, with clustering performance that does not improve with dimensionality reduction, indicating nonlinearly separable data. K-means clustering does respond to previously noted structural boundaries, such as the Pacific-North American plate boundary (PB), Southwest Deforming Zone (SDZ), and Central Deforming Zone (CDZ). HDBSCAN clustering produces more continuous clusters with fewer assumptions about the nature of the clusters. The clusters respond to previously noted structural boundaries, such as the PB, SDZ, and CDZ, as well as lithologic boundaries. Dimensionality reduction improves the continuity of the retrieved clusters, but removes fine-scale cluster separations within the sedimentary sequences and mixed lithologies within the fault-related damage zones.

Included in

Physics Commons