AI Unlocks Hidden Slow Slip Events on the San Andreas Fault: What DeepStrain Found

Most earthquakes announce themselves with a jolt. But beneath the Parkfield section of the San Andreas Fault in California, a quieter class of movement has been happening in plain sight—slow slip events (SSEs) that release energy over days or weeks without producing seismic shaking. Now, a deep-learning algorithm called DeepStrain has revealed that these silent events are far more common than previously thought, and that they play a direct role in triggering the small tremors known as low-frequency earthquakes (LFEs).

The Challenge of Detecting Silent Slip

Slow slip events are notoriously difficult to catch. Borehole strainmeters (BSMs)—instruments buried deep in cylindrical holes—can detect crustal deformation at the nanostrain level, making them sensitive enough to record transient strains from aseismic creep and episodic tremor and slip (ETS) . But the signals from small, short-lived SSEs are easily buried in noisy time-series data that human analysts often miss . GPS networks, the traditional tool for measuring crustal motion, are even less effective: the strain rates from these deep events are often below GPS detection thresholds .

How DeepStrain Works

Developed by geophysicist Zahra Zali at GFZ Potsdam, DeepStrain is a deep-learning model trained specifically on borehole strainmeter recordings from the NSF's Network of the Americas (NOTA) . The algorithm learns to recognize the subtle waveform patterns of slow slip within the high-dimensional noise of continuous strain data. The code and preprocessing pipeline were released publicly in August 2025, allowing other researchers to apply the method to other fault zones .

When applied to the Parkfield region, DeepStrain achieved a remarkable result: it detected 90% of previously manually cataloged SSEs and, more importantly, identified 21 new SSEs that had been missed by manual analysis . This roughly 30% increase in the known event catalog provides a far more complete picture of fault behavior in this intensively studied segment of the San Andreas.

The Temporal Link to Low-Frequency Earthquakes

Perhaps the most significant finding emerged when the team analyzed the timing of the newly detected SSEs relative to LFEs. The data showed that slow slip events were often followed by low-frequency earthquakes . This temporal sequence strongly suggests a causal mechanism: the aseismic slow slip loads or triggers the seismogenic patch that later generates the LFE.

This result is consistent with prior work showing that tremor and LFE activity near Parkfield share the same moment-duration scaling as slow slip events, implying they are physically linked . Low-frequency earthquakes have long been interpreted as seismic indicators of surrounding aseismic slip , but DeepStrain provides the clearest geodetic evidence yet that individual slow events precede and likely trigger those small earthquakes.

Implications for Fault Monitoring and AI in Geophysics

Expanded Detection Capability

DeepStrain demonstrates that AI can extract geodetic signals below the detection threshold of both GPS networks and manual strainmeter analysis. This expanded catalog of SSEs enables more robust statistical studies of fault behavior, recurrence intervals, and the conditions that lead to larger earthquakes .

A Causal Link Between Aseismic and Seismic Slip

The observation that SSEs systematically precede LFEs supports models in which slow slip loads nearby fault patches, potentially bringing them closer to failure. This has direct relevance for understanding earthquake nucleation and recurrence on the San Andreas Fault—a critical region for seismic hazard assessment .

Real-Time Operational Potential

Because DeepStrain can be deployed on continuous borehole strainmeter data, it offers a tool for near-real-time detection of transient deformation that might precede larger earthquakes. The NOTA network already maintains the necessary strainmeter infrastructure and makes both data and processing tools available to the research community . This could transform how earthquake early warning systems incorporate geodetic data.

Broader AI-in-Geophysics Precedent

This work joins a growing body of evidence that deep learning can systematically extract geophysical signals invisible to traditional methods. Similar approaches—such as CNNs for tremor detection in Cascadia and deep learning for LFE identification on the San Andreas—have shown that AI can serve as a "force multiplier" for existing monitoring networks . DeepStrain proves that the same principle applies to borehole strainmeter data, a key sensor type for detecting transient slip in the deep roots of faults.

Caveats and Open Questions

The precise architecture of DeepStrain (whether it uses a convolutional, recurrent, or transformer-based design) is not detailed in publicly available summaries. The full methodological details are in the Nature Communications paper (doi: 10.1038/s41467-026-74095-9) . Additionally, the algorithm has so far been validated only on the Parkfield segment; its performance on other fault zones with different strainmeter configurations and noise characteristics remains to be tested.