Southern California’s seismic network has put machine learning into the work that follows an earthquake. Its 2025 evaluation reported better automatic measurements and locations, with less subsequent analyst work.[1]
Why the arrival time matters
An earthquake sends different waves through the ground. P waves arrive before S waves, and the interval between them helps establish how far the source is from a recording station. Combining observations from several stations narrows down its position. A seismogram therefore contains more than evidence that something shook: the timing inside the trace helps reveal where it happened.[2]
The instruments record motion in three directions: north–south, east–west and vertically. Those channels provide complementary views of the same disturbance. One earthquake can produce many useful observations across a network, so an increase in the number of measured wave arrivals is different from an increase in the number of earthquakes.[2]
Learning to separate waves from noise
PhaseNet was trained using recordings and arrival times labelled by experienced Northern California analysts. It receives the three waveform components together and estimates where P arrivals, S arrivals and background noise occur. Peaks in its output mark candidate arrival times. Its learned patterns offer a richer way to interpret a trace than looking only for a sudden increase in signal energy.[3]
Three parts of an arrival-time measurement
| Part | What it contributes |
|---|---|
| Recorded waveform | The changing ground motion that the model examines.Source 3 |
| P and S classifications | The type of seismic wave associated with a candidate arrival.Source 3 |
| Arrival-time pick | A time marker that can be compared across stations.Source 3 |
S waves are particularly valuable because their later arrival can be difficult to separate from the lingering P-wave signal. Better S picks add information that helps constrain earthquake depth as well as location.[3]
What changed in operational use
Southern California’s reported results
| Measure | Value | Context |
|---|---|---|
| Wave arrivals | About 2–3 times as many automatic picks, especially S waves. | |
| Timed review | About 30 seconds faster per event on average. | One analyst reviewed 32 pairs of events matched for location and magnitude. |
| Backlog test | 476 of 1,064 events could be finalised without manual adjustment after processing. |
Joining separate readings into an event
A further challenge comes when many earthquakes occur close together. Finding a plausible arrival at one station does not settle which event produced it. GaMMA, a separate machine-learning method, groups readings that can be explained by a common earthquake. Its model considers both arrival time and signal amplitude alongside the distances between a proposed source and the stations.[4]
This is a physically constrained grouping problem. Waves take time to travel, and their amplitudes generally decline with distance. A proposed group should fit those relationships. GaMMA updates the grouping and the estimated source together, rather than simply treating every nearby spike as part of the same event. More precise catalogues can then help researchers examine earthquake sequences and the structures on which they occur.[4]
The Southern California team incorporated GaMMA into a pipeline for readings that its initial system had left unassociated. When that pipeline finds a candidate event, it enters the event database and passes through another processing stage that adds readings from stations outside the original local group. The result reaches analyst review with more of the network’s evidence attached.[1]
Sources & further reading
- Improvements from incorporating machine learning algorithms into near real-time operational post-processing Scientific Reports · 7 Aug 2025 · Accessed 11 Oct 2026
- Seismographs — Keeping Track of Earthquakes U.S. Geological Survey · Accessed 11 Oct 2026
- PhaseNet: A Deep-Neural-Network-Based Seismic Arrival Time Picking Method Zhu and Beroza / arXiv · 8 Mar 2018 · Accessed 11 Oct 2026
- Earthquake Phase Association Using a Bayesian Gaussian Mixture Model Journal of Geophysical Research: Solid Earth · 29 Mar 2022 · Accessed 11 Oct 2026