[Paper Review] Breaking Cascadia's Silence: Machine Learning Reveals the Constant Chatter of the Megathrust

Bertrand Rouet‐Leduc, Claudia Hulbert+1 more|arXiv (Cornell University)|May 17, 2018

earthquake and tectonic studies28 references4 citations

TL;DR

This study uses supervised machine learning to extract continuous tremor-like seismic signals from raw data across Vancouver Island, revealing that the Cascadia megathrust emits a persistent, low-amplitude seismic signal that precisely tracks the fault's displacement rate throughout the slow slip cycle. The signal accounts for up to 300 times more energy than cataloged tremor events and provides real-time, high-resolution insight into fault slip dynamics, offering a new indirect probe of megathrust physics and potential earthquake precursors.

ABSTRACT

Tectonic faults slip in various manners, ranging from ordinary earthquakes to slow slip events to aseismic fault creep. The frequent occurrence of slow earthquakes and their sensitivity to stress make them a promising probe of the neighboring locked zone where megaquakes take place. This relationship, however, remains poorly understood. We show that the Cascadia megathrust is continuously broadcasting a tremor-like signal that precisely informs of fault displacement rate throughout the slow slip cycle. We posit that this signal provides indirect, real-time access to physical properties of the megathrust and may ultimately reveal a connection between slow slip and megaquakes.

Motivation & Objective

• To identify and quantify a continuous seismic signal in the Cascadia subduction zone that may account for missing energy in slow slip events.
• To develop a machine learning framework that uses continuous seismic data to estimate fault displacement rates with higher temporal resolution than GPS.
• To investigate whether the persistent seismic signal originates from the same physical processes as known tremor bursts.
• To assess whether this signal can serve as a real-time proxy for fault slip dynamics, improving monitoring of stress transfer to the locked zone.
• To explore the scaling of slow slip physics from laboratory experiments to natural tectonic systems.

Proposed method

• Supervised machine learning is applied to 40 Hz continuous seismic data from the Canadian National Seismograph Network (CNSN), using GPS-derived displacement rates as the labeled target.
• Statistical features—amplitude, frequency, and energy characteristics—are extracted from hourly to daily time windows of seismic data.
• A random forest regression model is trained to map seismic features to GPS displacement rates, with model performance validated on independent test data.
• The model is tested on data from 2009–2012 for training and 2013–2017 for testing, using a fixed window length of 1–60 days.
• The method compares predictions from full continuous data against predictions restricted to cataloged tremor events to isolate the contribution of background tremor.
• Energy content of the continuous signal is quantified and compared to the cumulative energy of cataloged tremor events to assess missing energy.

Experimental results

Research questions

• RQ1 Can continuous seismic data reveal fault displacement rates with higher temporal resolution than GPS?
• RQ2 Is there a persistent, low-amplitude seismic signal in the Cascadia megathrust that correlates with slow slip rates?
• RQ3 Does the energy of this continuous signal account for the discrepancy between GPS-measured displacement and cataloged tremor energy?
• RQ4 Are the physical mechanisms behind the continuous signal and classical tremor events the same?
• RQ5 Can machine learning extract meaningful geophysical signals from raw seismic data that are missed by traditional event-based catalogs?

Key findings

• The continuous seismic signal tracks the fault's displacement rate with high fidelity on an hourly basis, outperforming GPS, which is too noisy for sub-daily resolution.
• The signal accounts for approximately 300 times more seismic energy than cataloged tremor events, resolving a long-standing discrepancy in energy budget calculations.
• The normalized energy of the continuous signal correlates strongly with cataloged tremor energy during peak slow slip, suggesting a common physical origin.
• The signal persists even during non-tremor periods, indicating that tremor is not episodic but continuous or near-continuous, challenging previous assumptions.
• The signal's statistical fingerprint matches laboratory observations of slow slip, suggesting that the underlying frictional physics may scale from lab to Earth.
• The model trained on continuous data predicts displacement rates more accurately than models restricted to cataloged tremor events, demonstrating that non-event data contain critical information.