How a telecom fibre became part of an earthquake monitoring network

A telecom cable near Ridgecrest, California, has supplied earthquake observations to Southern California’s seismic network.[2]

Distributed acoustic sensing sends laser pulses down an optical fibre and measures changes in the returning light. Tiny deformations along the cable become a record of strain. Thousands of positions can supply measurements without thousands of separately installed seismometers. But roadside fibre also feels vehicles, and its contact with the ground varies along the route.[3]

More measurements are useful only if their relationships survive the processing. Treat every position as an isolated alarm and the system loses much of what makes a long cable valuable: it can show a disturbance travelling through space.

Read broadly, report selectively

The team’s reported Ridgecrest configuration had 10,000 channels, of which bandwidth allowed 5,000 to stream to Caltech. PhaseNet-DAS examined those streamed channels together. For integration with the ANSS Quake Monitoring Software, the researchers selected 18 positions roughly five kilometres apart. Their May 2025 preprint reported eight already contributing to network operations; the study subsequently appeared in Seismological Research Letters in 2026.[1][2]

The two channel counts answer different questions. A broad input helps the model recognise a wavefront. A smaller output supplies the established network with useful, spaced observations. Reducing what enters the operational system therefore need not mean depriving the AI of the surrounding evidence.

PhaseNet-DAS is designed to read a two-dimensional pattern: time on one axis and position along the fibre on the other. Traffic tends to produce a local disturbance; an earthquake can register coherently across a much longer section. The model uses that context to locate the arrivals of P and S waves, whose timing helps locate an earthquake.[3]

Its training also used spatial consistency. An earlier seismic model proposed candidate labels, then a physical association method filtered them before they trained the fibre-specific network. The resulting model was developed on four California arrays; performance at a different installation still depends on its recording conditions.[3]

The route into a working network

In the integration tests, the model processed a rolling 70-second window in approximately one second per picking task. That is inference time, not a measured one-second earthquake warning. For a Lamont earthquake sequence, the workflow detected more than 400 aftershocks; a conventional threshold method missed many and produced traffic-related picks.[1]

A rolling window provides recent history each time the model examines the newest data. Keeping that history is different from waiting for a fresh, non-overlapping block to finish. The relevant engineering question is whether processing can keep up as new observations arrive.

The software translates selected results into the formats used by the existing network. The described system excludes amplitude information while cable response and ground coupling require further calibration.[1]

Arrival timing and amplitude have different jobs. A wave’s arrival helps constrain where an event occurred; its size carries additional information that depends on how faithfully the instrument represents ground motion. Adding one reliable measurement need not wait for every other measurement to be equally mature.

That is the useful design choice here: preserve the cable’s spatial detail where it helps interpretation, then pass an intelligible result to the people and software already monitoring the region. An enormous sensor becomes useful through a carefully chosen connection.

Report an error or suggest a correction