Researchers working at Los Alamos National Laboratory applied deep learning to seismic data and learned that tremor and slip occur at all times — before and after known large-scale slow-slip earthquakes — rather than intermittently in discrete bursts, as previously believed.

“The work tells us that the physics of friction on faults appears to have universal characteristics — something we suspected but could not prove,” said Bertrand Rouet-Leduc, a geophysicist in the Geophysics group at the Energy Department's laboratory.

Map showing location of the Cascadia Megathrust Fault in the Pacific Northwest. Source: University of OregonIn the research, the team trained what is known as a "convolutional neural network" — a form of deep learning — on a tremor catalog created at the University of Washington. The catalog uses several years of seismograms from one seismic station on Vancouver Island in the Cascadia region of the Pacific Northwest. Tremor events that were initially identified by multi-station methods formed the training set. The team then used the deep learning model to find more events.

The researchers said that slow earthquakes, which are known to sometimes precede major earthquakes on continental faults and in subduction zones, build stress cyclically. They may trigger large earthquakes on neighboring, highly stressed locked faults. Yet even in Cascadia, a clear case of a highly stressed fault, previous research had only observed intermittent and discrete slow earthquakes and their associated tremors.

The research team discovered that the neural network provided a continuous measure of tremors and placed clearer time bounds on slow-slip events than previous methods had established. They said the neural network identified faint tremor signals months before traditional methods detected slow slip by measuring sometimes tiny elevation changes in the landscape.

“The study suggests that slow slip within megathrust zones is not random. This research is part of a body of work that tells us that precursors are taking place much more frequently than previously thought, in agreement with laboratory experiments and theory,” said Rouet-Leduc.

The paper, “Probing Slow Earthquakes with Deep Learning,” was published in Geophysical Research Letters. This work is part of ongoing research at Los Alamos that has identified a continuous acoustic signal emitted by slow-slip events. The signal can be read at any instant to indicate the time remaining to fault failure in both laboratory and real-world quakes.