SURF: An Automated Method for Building Nonplanar 3D Fault Models from Earthquake Hypocenters

Abstract Accurately characterizing 3D fault geometry is vital for improving our understanding of earthquake behavior and informing the development of seismic hazard models. Despite their importance, subsurface fault structures tend to be poorly constrained because of limitations in observational data. Improvements to the seismic networks and earthquake detection algorithms have increased the precision and volume of earthquake catalogs, which help illuminate detailed subsurface fault structure and provide the most direct information available about fault geometries at depth. We present a Python package to automate generating 3D fault geometries directly from hypocentral seismicity patterns. This method begins with clustering events based on their spatial density, identifying coherent patterns. Nearby clusters are then merged based on the similarity of their orientations. We fit nonplanar surfaces using support vector regression to balance surface accuracy with minimal deviations from planarity. The fault models are output as quadrilateral meshes at user-defined resolution. In the process of generating the 3D fault surfaces, we compute the spatial density of seismicity around the surface and the planarity as quantitative metrics of the model outputs. As a proof of concept, we apply this approach to the San Andreas–Calaveras fault junction region and the 2019 Ridgecrest earthquake sequence, both in California, which contain complex subparallel faults well defined at the Earth’s surface and abundant microseismicity. These case studies demonstrate the method’s ability to model complex fault structures, including long continuous fault surfaces, crossing faults, variably dipping segments, and subparallel faults. We test the method on both standard network catalogs and double-difference relocated catalogs. We find that our seismicity-based fault model results align with published 3D models that incorporate additional constraints and interpretations (Plesch et al., 2020; Aagaard and Hirakawa, 2021). This workflow provides a low-user-input solution for estimating fault geometries at depth from earthquake catalogs.