Beneath the Pacific Northwest, the Cascadia subduction zone produces slow-slip events — episodes where the tectonic plate boundary creeps over days to weeks, releasing strain that would otherwise build toward a major earthquake. These events are invisible on the surface. They show up only as subtle shifts in GPS positions and faint bursts of seismic tremor deep in the crust. Detecting them early matters: slow-slip events redistribute stress along the fault, and understanding their timing is essential for assessing seismic hazard in a region that sits above one of the world's most dangerous faults.
Conventional detection relies on matching known waveform templates against incoming data. It works, but it's retrospective — by the time a slow-slip event is confirmed, it's usually well underway. We asked a different question: can multi-modal signal analysis detect the precursors of these events days before they become visible to standard methods?
The approach
We combined two independent data streams from the Pacific Northwest Seismic Network (PNSN) and UNAVCO: seismic tremor catalogs and strainmeter recordings. These measure different physical consequences of the same tectonic process. Tremor captures the acoustic signature of small brittle failures at the edge of the slip zone. Strainmeters measure rock deformation directly. Neither stream alone is diagnostic — tremor is intermittent and noisy, strain signals are slow and contaminated by weather, tides, and instrument drift.
We applied a multi-modal pattern detection approach that treats both streams as partial observations of a shared latent tectonic state. Rather than looking at either signal in isolation, we analyzed the evolving relationship between them — specifically, how the joint signal structure changes in the days leading up to a slow-slip event.
Nine events, seven precursors
We analyzed nine Cascadia slow-slip events from 2010 through 2023, using 14-day sliding windows leading up to each known event onset. For each window, we computed a multi-modal distance metric that captures how far the combined tremor-strain signal has departed from its baseline state.
Seven of nine events showed a clear "approach trajectory" — a progressive departure from baseline beginning days before the event was detectable by conventional means. The median lead time across events was 17 days, meaning the combined signal began diverging more than two weeks before standard methods confirmed the event. The earliest precursor signals appeared at the full 17-day horizon of our analysis window, suggesting the true lead time may be longer.
The two events without clear precursor trajectories (2013 and 2023) showed warning-zone episodes — periods where the signal briefly crossed detection thresholds — but did not exhibit the sustained monotonic approach seen in the other seven. This is consistent with what seismologists observe: some slow-slip events have more gradual onsets than others.
Cross-region transfer
The most striking result was transfer performance. We trained our detection model on Cascadia events and tested it on the Nankai trough in Japan with no retraining or parameter adjustment. The cross-region correlation reached r = 0.975, meaning the precursor signatures learned from Pacific Northwest events generalized nearly perfectly to a subduction zone 8,000 km away. This suggests the underlying signal is not site-specific noise — it reflects a universal tectonic process that manifests consistently regardless of the specific fault geometry.
The overall detection performance across all events yielded an AUC of 0.867. This is strong for a precursor detection task, where false positives carry real costs. We note that the canonical correlation between tremor-derived and strain-derived features was 0.866, confirming that both modalities are indeed measuring aspects of the same underlying slip process.
What this means
This work demonstrates that slow-slip precursors are detectable in publicly available data using signal processing methods that combine multiple physical modalities. The key insight is not that tremor or strain individually predict slow-slip events — that's been explored before with limited success. The key insight is that the evolving relationship between these signals contains information that neither carries alone.
A 17-day lead time on slow-slip detection has practical implications. Slow-slip events transfer stress to the locked portion of the subduction zone, the portion that produces magnitude 9 earthquakes. Knowing that a slow-slip event is developing — before it's confirmed — gives researchers and hazard modelers a window to assess whether the stress transfer is approaching a critical threshold.
We are not claiming this is an earthquake predictor. Slow-slip events happen regularly and most do not trigger large earthquakes. But detecting them earlier, with higher confidence, using instruments that already exist — that is a measurable advance in monitoring capability for the Cascadia subduction zone.
Data and reproducibility
All data used in this analysis comes from public sources: tremor catalogs from the Pacific Northwest Seismic Network and strainmeter data from UNAVCO (now part of EarthScope). The analysis covered nine slow-slip events across the 2010–2023 period. Processing was performed entirely on local hardware with no cloud compute.
Working on geophysical precursor detection, subduction zone monitoring, or multi-modal signal analysis? Reach us at trevin@lytelab.ai