FaultNet: Deep Learning Reconstruction of Three-Dimensional Fault Geometry from Earthquake Hypocenters

Yanlan H. Hu, & Gregory C. Beroza

Submitted August 30, 2026, SCEC Contribution #15567, 2026 SCEC Annual Meeting Poster #153

Accurately characterizing fault geometry is fundamental to understanding earthquake processes, stress transfer, and dynamic rupture, as well as assessing seismic hazard. Recent advances in earthquake detection, phase picking, and relocation have produced high-resolution earthquake catalogs that form dense three-dimensional point clouds delineating active fault systems; however, automatically extracting individual fault structures from these catalogs remains challenging because of complex fault geometries, intersecting faults, and heterogeneous seismicity distributions.

This motivated us to develop FaultNet, which is a deep learning framework that reconstructs three-dimensional fault geometry directly from earthquake hypocenters by formulating the problem as semantic segmentation of point clouds. FaultNet employs two PointNet++ networks to identify fault-related earthquakes and learn geometric feature embeddings that capture multi-scale structural characteristics. These embeddings are integrated across overlapping sliding windows to group earthquakes into fault clusters, followed by reconstruction of continuous three-dimensional fault surfaces.

We apply FaultNet to earthquake catalogs spanning induced seismicity, mainshock-aftershock sequence, volcanic seismicity, and California regions where fault geometries are incompletely resolved by the Community Fault Model. The reconstructed fault geometries are evaluated using three defined quantitative metrics and independent seismological observations, demonstrating that FaultNet accurately recovers complex fault structures, including intersecting and heterogeneous faults. By directly learning geometric patterns from earthquake distributions, FaultNet provides an automated and generalizable framework for reconstructing active fault systems from increasingly large and detailed earthquake catalogs, offering new opportunities for investigating fault architecture and earthquake processes.

Citation
Hu, Y. H., & Beroza, G. C. (2026, 08). FaultNet: Deep Learning Reconstruction of Three-Dimensional Fault Geometry from Earthquake Hypocenters. Poster Presentation at 2026 SCEC Annual Meeting.

Related Projects & Working Groups
Seismology