AI reveals hidden fault movements beneath the San Andreas Fault

Working with one of the world’s most closely monitored fault systems, researchers from the GFZ Helmholtz Centre for Geosciences, EarthScope and Stanford University built the most comprehensive catalog yet of these elusive events. Fourteen of the newly discovered slow slip events were independently confirmed by nearby creepmeters, strengthening evidence that the signals represent genuine aseismic fault movement.

The discovery fills a long-standing observational gap between earthquakes, which release energy within seconds, and larger slow-slip episodes that can unfold over days or weeks. Instead, these newly cataloged events typically lasted 25–100 minutes, revealing a previously hidden form of fault movement that conventional monitoring methods had struggled to detect.

Parkfield has long occupied a unique place in earthquake science. The central California segment of the San Andreas Fault is one of the most heavily instrumented fault systems on Earth, making it a natural laboratory for studying how tectonic stress accumulates and is released. Despite decades of continuous monitoring, the smallest transient slip events remained buried within enormous volumes of noisy geophysical data.

To uncover those hidden signals, the team analyzed eight years of continuous borehole strainmeter observations using an unsupervised deep-learning workflow. Instead of searching for predefined patterns, the algorithm learned to recognize recurring deformation signatures on its own, separating likely tectonic events from environmental and instrumental noise.

“Faults can move in ways that do not generate strong seismic waves and therefore escape traditional earthquake detection methods. We wanted to know whether important fault slip processes might be hidden within years of continuous deformation measurements,” said lead author Dr. Zahra Zali.

The automated workflow recovered about 90% of the slow-slip events listed in an existing manually compiled catalog while identifying 21 previously unknown events.

Fourteen of those newly recognized episodes were also recorded by nearby creepmeters, strengthening the conclusion that the signals reflected genuine aseismic fault slip rather than instrumental artifacts.

After identifying the events, the researchers calculated where the slips most likely originated. Their analysis indicates the transient deformation occurred at shallow depths, generally less than 4 km (2.4 miles), and was predominantly consistent with the right-lateral motion that characterizes the San Andreas Fault.

These quiet fault movements don’t appear to happen in isolation.

When the researchers overlaid their new slow-slip catalog with records of low-frequency earthquakes (LFEs), a clear temporal pattern emerged. LFE activity consistently increased after the occurrence of slow-slip events, with the strongest relationship observed for earthquakes occurring within 10 km (6.2 miles) of the reference strainmeter and at depths shallower than 20 km (12.4 miles). The pattern suggests that transient aseismic slip can modify stress conditions within the fault system and influence subsequent low-frequency seismic activity.

“Our results show that these slow fault movements are not isolated phenomena. They appear to be linked to changes in seismic activity, which suggests that slow slip may play an important role in how stress evolves along active faults,” said Prof. Patricia Martínez-Garzón of the GFZ Helmholtz Centre for Geosciences.

The AI-driven catalog also challenges a long-standing assumption about the physics of slow-slip events.

For years, many studies have described slow-slip events using a linear moment-duration relationship. The Parkfield observations instead point to an approximately cubic moment-duration scaling law, placing these short-duration events much closer to the scaling behavior observed for ordinary earthquakes. While the researchers note that additional observations will help better constrain the geometry of these events, the findings add new evidence that at least some forms of aseismic fault slip may follow earthquake-like physical scaling.

Without machine learning, many of these events would likely have remained hidden within years of continuous strain measurements. As geophysical monitoring networks continue to expand, researchers increasingly face datasets too large for manual analysis. The Parkfield study demonstrates how AI can sift through continuous observations, identify subtle deformation patterns and reveal tectonic behavior that conventional workflows can miss.

The researchers stress that the newly detected events are not earthquakes and should not be interpreted as evidence that a damaging earthquake is imminent. Instead, they provide a clearer view of how stress is redistributed between major earthquakes. By illuminating this previously hidden part of the fault-slip spectrum, the work offers scientists another tool for understanding how active faults evolve through time and how silent deformation fits into the broader earthquake cycle.

References:

¹ Zali, Z., Martínez-Garzón, P., Mencin, D. et al. Slow slip modulates low-frequency seismicity on the Parkfield segment of the San Andreas Fault. Nat Commun 17, 5137 (2026). https://doi.org/10.1038/s41467-026-74095-9

² San Andreas Fault: Hidden movements revealed by artificial intelligence – GFZ – July 9, 2026