Weiqiang Zhu ("Wei-Chiang Zhu")
Advances in seismology have long followed the growth of seismic observations. Seismic networks have accumulated petabytes of continuous waveforms and decades of earthquake catalogs, while new sensing technologies such as distributed acoustic sensing (DAS) turn telecommunication fibers into arrays of thousands of channels recording terabytes per day. Analyzing these datasets accurately and at scale is becoming a challenge. My group develops deep-learning and machine-learning methods spanning earthquake detection, inversion, and forecasting.
For detection, PhaseNet picks P/S arrivals from continuous waveforms; PhaseNet+ incorporates multi-task learning, jointly picking arrival times and first-motion polarities and estimating origin times; and EQNet2 incorporates multi-station input, recovering weak coherent signals across stations and suppressing noise-induced false detections at single stations. DAS extends this coherence to channels spaced meters apart, so we formulate phase picking as semantic segmentation (PhaseNet-DAS) and event detection as object detection (DASNet), tracking arrivals across adjacent channels rather than picking each independently. The phases picked by these models are then associated into events by unsupervised clustering (GaMMA, VORA) and inverted for hypocenters (ADLoc) and subsurface velocity structure (ADTomo) within an automatic-differentiation framework. Integrating these steps into the cloud-native workflow QuakeFlow, we build enhanced earthquake catalogs in regions such as the Mendocino Triple Junction and across California, containing up to an order of magnitude more events than standard catalogs, which support studies of earthquake physics, fault structure, and seismic hazard.
Building on such catalogs, a masked Transformer (QuakeFormer) learns site- and path-specific effects directly from growing archives of recorded ground motions, and a diffusion model (QuakeGen) generates evolving fields of aftershock rate in space and time, capturing fault-controlled patterns. On the physics side, we simulate earthquake sequences and aseismic slip with coupled fluid flow, permeability evolution, and rate-and-state friction to quantify fault valving, in which fluid overpressure builds during the interseismic period as healing seals the fault and is released when earthquakes enhance permeability, driving fluid-driven aseismic slip and earthquake swarms.
These methods turn expanding seismic datasets into more complete catalogs, sharper images of fault and velocity structure, and better forecasts of aftershocks and ground motion, which together advance the study of earthquakes and the hazards they pose. A full list of publications and open-source code, models, and datasets is on the AI4EPS group website.
