FaultNet: Inferring Complex Fault Structures from Earthquake Hypocenter Distributions

Accurately characterizing fault geometry is fundamental to understanding earthquake processes, including stress transfer and rupture dynamics, as well as for assessing seismic hazard. Recent advances in earthquake detection, phase picking, and relocation have produced high-resolution hypocenter catalogs that can be viewed as three-dimensional point clouds recording the geometry of active fault systems. Extracting individual fault structures from these catalogs remains challenging due to complex fault geometries and heterogeneous seismicity distributions. Here, we present FaultNet, a deep learning framework that formulates fault extraction from earthquake hypocenters as a three-dimensional semantic segmentation problem. Built upon PointNet++, the denoising and feature-learning models in FaultNet hierarchically process point coordinates to learn feature embeddings and are trained on synthetic earthquake catalogs. These feature embeddings are used to identify individual fault clusters and reconstruct continuous three-dimensional fault surfaces. We apply FaultNet to three catalogs representing induced seismicity, mainshock-aftershock sequences, and volcanic seismicity. Across these diverse settings, the method successfully resolves coherent fault structures, including intersecting and segmented faults. The reconstructs fault geometries perform well across our evaluation metrics and are consistent with independent geological and seismological observations. Our results demonstrate that deep learning provides an effective and generalizable approach for reconstructing complex fault systems directly from earthquake hypocenter distributions.