Along California's San Andreas Fault, the North American and Pacific plates grind slowly past one another. Part of that motion happens in silence: in so-called slow-slip events, rock shifts over the course of hours without any perceptible shaking. Such movements release stress underground – but because their signals are so faint, they had largely eluded science.
A team led by Zahra Zali of the GFZ Helmholtz Centre for Geosciences in Potsdam has now made these silent motions visible. Working with colleagues from Stanford University and the research infrastructure EarthScope, the researchers developed an AI method that filtered dozens of short slow-slip events out of years of measurements. The study appears in the journal Nature Communications.
A learning program finds what instruments miss
The work drew on data from borehole strainmeters at the Parkfield section of the fault – one of the most densely monitored earthquake zones in the world. These highly sensitive instruments register tiny deformations in the Earth's crust, but they generate enormous volumes of data in which individual weak signals easily disappear.
Rather than searching for predefined patterns, the program learned directly from the continuous measurements and grouped similar deformation patterns together. "Artificial intelligence allowed us to recognize patterns that would otherwise have gone unnoticed," Zali explains. The result was the first catalogue of short slow-slip events for Parkfield.
One relationship stood out: the silent shifts were systematically followed by increased activity of so-called low-frequency earthquakes. The study thus provides evidence that slow-slip processes help shape the subsequent course of seismic activity – a building block that could, in the long run, help researchers better understand how major faults behave.
For regions along such plate boundaries, every additional insight into the hidden part of the earthquake cycle is valuable. The method also shows how decades of accumulated measurements can be re-examined with learning algorithms – surfacing processes that earlier instruments had missed.