PhaseNet: How AI Marks the Waves That Help Locate an Earthquake

Picture three seismic stations recording a busy afternoon. Each draws a continuous trace. A sharp movement at one station could be the first arrival from an earthquake. A second mark at a nearby station might belong to that same event, or to another earthquake that happened close in time. Before anyone can put a dot on a map, somebody has to work out when the waves arrived and which arrivals belong together.

PhaseNet helps with the first part. Its output is useful precisely because it is narrower than the phrase ‘AI detects earthquakes’ suggests. It marks plausible arrivals on a station’s recording. Turning those marks into an earthquake catalogue requires a separate judgement about how signals fit together across the network. That distinction explains both the promise of the technology and the limits of a good-looking result.

What PhaseNet actually marks

In their paper in the January 2019 issue of Geophysical Journal International, Weiqiang Zhu and Gregory Beroza describe a neural network trained using analyst-labelled arrivals from Northern California. It examines waveform data recorded in three directions and outputs scores over time for P waves, S waves and noise. A peak in the appropriate score can be used to select an arrival time, a step seismologists call picking.

A pick is an estimate of when a type of wave reached a particular sensor. P waves typically arrive before S waves from the same earthquake at the same station. The time between their arrivals can help constrain distance, given assumptions about how fast the waves travel through the ground. One station’s pair still leaves open where around the station the source might be. Other stations add constraints through their locations and the times they recorded.

The model learns from examples already interpreted by analysts. That makes the training labels valuable, but it also means readers should ask what sort of recordings the evaluation covered. Performance on labelled examples from one network gives evidence about that setting. Whether the same approach handles a different mix of instruments, noise and geology is a question to check with local recordings.

Which arrivals belong to the same event?

Imagine two small earthquakes occur close together in time. At one station, the first earthquake supplies a clear P arrival, while the second supplies the strongest S arrival in the window. Pair those two because they are easy to see and the resulting time gap can look perfectly plausible. The mistake only becomes apparent when you ask whether other stations recorded arrivals that fit the same proposed source.

A separate 2022 study on GaMMA by Zhu and colleagues describes one way to make that comparison. The method groups arrivals using their timing and amplitude together with station locations, while estimating properties of the sources. The researchers tried it on synthetic data and on the 2019 Ridgecrest earthquake sequence, using PhaseNet picks in the latter case. Those evaluations address a different question from how closely a pick matches an analyst’s mark: whether a set of picks can be assembled into a plausible event.

The distinction also shows up in a 2024 study of rapid aftershock analysis, where the researchers used PhaseNet to pick arrivals and GaMMA to group them and obtain preliminary event estimates. The word ‘preliminary’ matters. A useful early result can guide investigation while the team continues to assess what the recordings support. These studies concern seismic signals that have already reached instruments; they do not demonstrate prediction of an earthquake before it starts.

What to ask when an AI catalogue grows

Suppose a research team runs a new picking and association workflow on recordings it has already analysed and reports more events. That could mean it recovered small earthquakes the earlier catalogue missed. It could also reflect a change in which signals were accepted, or how the workflow split overlapping arrivals. To distinguish those explanations, start with a sample of the events that appear only in the new catalogue. Ask whether the supporting recordings tell a coherent story across stations and whether reviewers can reproduce the judgement.

• Ask what the benchmark counts. A score for arrival timing, a score for grouping picks and an error estimate for source location answer different questions.
• Look at hard cases. Ask how the workflow handles overlapping events, noisy stations and signals that are visible at only part of the network.
• Keep the chain of evidence. Link catalogue entries back to the station recordings, the picks and the settings used to make them. That makes disagreements easier to investigate.

A related problem appears in our article on AI-suggested fresco joins. A candidate match gives researchers a place to look; they still need evidence that the fragments fit. The analogy has a limit: seismic arrivals are indirect observations and the proposed source is inferred from how waves travel. But the reading habit is the same. Find out what the model marked, then ask what other evidence makes the larger claim credible.

For readers outside seismology, that habit is the most useful takeaway. If a system claims to have found more of something, look for the step that turns a local signal into a countable event. That step may involve a second model, physical constraints or expert review. A tally is much easier to interpret when you know how the individual entries earned their place.

Sources

• Zhu and Beroza, PhaseNet: seismic arrival-time picking (2019 issue).
• Zhu and colleagues, Earthquake phase association using a Bayesian Gaussian mixture model (2022).
• Rapid analysis of aftershock processes with machine learning methods, Geophysical Journal International (2024).