From Phase Pickers to Seismic Foundation Models

Picture an earthquake-monitoring pipeline where every job runs its own AI model: one to detect an event, one to pick the P arrival, another for the S arrival, another to estimate magnitude, another to read first motion. That is more or less how modern seismic machine learning grew up — a toolbox of narrow specialists, each trained from scratch on its own carefully labeled dataset.

Foundation models propose a different starting point: first learn one reusable representation of waveforms, then adapt it to each task with a much smaller labeled dataset. The interesting question isn’t whether seismic data are literally a language. It’s whether their recurring structure — noise, onsets, phases, amplitudes, frequency content, station quirks — is regular enough that a model can learn useful context before we ever tell it what a given trace means.

From rules to learned features

The through-line of this history is a single question: who decides what matters in a noisy trace? STA/LTA compares the energy in a short window against a long one; template matching asks whether a new trace resembles an event we have already seen. In both, a human hard-codes the signal to look for.

Deep networks handed that choice to the training data instead. Convolutional pickers learned the relevant features from labeled examples; attention-based models could weigh a wider stretch of the trace at once. Neither replaced seismology — they offered a new way to turn decades of waveform archives into predictions.

Why PhaseNet was a breakthrough—and why it is not a foundation model

PhaseNet struck a simple supervised-learning bargain: pair three-component waveforms with analyst-labeled P and S arrivals, then train a network to output P, S, and noise probabilities through time. Wherever a probability curve peaks, you have a pick. Its original evaluation showed just how much learned features could beat hand-designed picking rules [1].

That is exactly why PhaseNet caught on: it turns a slow, subjective judgment into a fast, repeatable mapping from waveform to probabilities. Earthquake Transformer pushed the idea further, learning event detection and P/S picking in one network [2], and SeisBench then packaged datasets, models, and a shared interface so they were finally easy to compare and reuse [3].

Their limitation isn’t that they are “bad at learning.” It’s that the training objective fixes the answer they are allowed to give from day one. A picker is optimized to pick; a magnitude regressor is optimized to estimate magnitude; neither knows how to do the other’s job. Move to a new region, a new sensor type, or a new task, and you are often back to collecting labels and making training decisions from scratch.

What does “foundation model” mean here?

Think of it like learning a language before choosing a profession. You pick up general vocabulary and how sentences fit together first; the specialized training — law, medicine, seismology — comes later. Knowing the language is not enough to practice law, but it makes the professional training far quicker.

For waveforms, that shared “vocabulary” is an embedding: a compact numerical fingerprint of each trace. Similar waveforms land near each other, so noise and clear arrivals pull apart into different regions of the space. Even an ordinary supervised picker builds such a space inside itself — the figure below projects PhaseNet’s internal embedding of real waveform windows, and noise, P, and S arrivals separate cleanly even though the model was never told which point is which. A foundation model sets out to learn exactly this kind of organized representation from broad, mostly unlabeled data, so it can later be adapted to more than one task [10].

Learning from data that have no labels

Unlabeled does not mean uninformative. Each sample constrains its neighbors, and a physically real signal has to fit smoothly into what surrounds it. Self-supervised learning turns that built-in structure into the training target itself — the data supervises the model, and no analyst has to annotate every trace.

Three recipes dominate. In masked prediction, part of the trace is hidden and the model must infer the missing piece from its surroundings. Contrastive learning pulls two views of the same signal together in the representation space and pushes unrelated ones apart. Reconstruction corrupts a waveform and trains the model to rebuild it. All three chase the same prize: a representation that keeps the signal structure worth reusing later.

Two recent waveform-model experiments

SeisLM is a self-supervised, encoder-only transformer trained on single-station waveforms [6]. It learns from masked latent features and a contrastive objective, then bolts on small task heads for event detection, phase identification, onset regression, and foreshock/aftershock classification. The takeaway: one representation, learned with no task labels at all, can be put to work on several different questions afterward.

U-Trans takes the reconstruction route instead [7]. Its U-Net/transformer hybrid is pretrained to rebuild waveforms that have been corrupted in both the time and frequency domains; its latent features are then fed — alongside the raw waveform — into heads for phase picking, location, magnitude estimation, and polarity classification. Both models are promising demonstrations that transfer works, not a claim that one frozen model is ready to drop into every network.

Transfer is the practical test

Say a model is pretrained mostly on California earthquakes and then sent to Taiwan, Japan, or an induced-seismicity field. The waveform “grammar” still rhymes, but the distribution shifts underneath it: different geology bends the propagation and reshapes the noise, station responses and orientations differ, sampling rates vary, and the events themselves may be smaller, shallower, or more tightly clustered.

Pretraining can make adaptation far more data-efficient, but it does not make that domain shift vanish. The honest workflow is unglamorous: hold out representative local data, compare against a local baseline, break the errors down by signal-to-noise ratio and distance, and report uncertainty instead of a single headline score. Systematic picker evaluations show plainly that transfer success depends on both the data domain and the task [4].

What could change—and what remains hard

If reusable waveform backbones mature, they could spare us from retraining the same wheel across detection, picking, source characterization, ambient-noise analysis, microseismic and volcano monitoring, and eventually new data types like DAS. Some groups are chasing more speculative routes: SeisMoLLM transplants a pretrained language-model architecture into seismic monitoring [8], while another line tests whether neural audio-compression autoencoders can squeeze and rebuild seismic waveforms well enough to double as a general-purpose representation [9]. These are live research directions, not yet an operational standard.

The hard problems matter just as much: curating geographically diverse data, controlling compute cost, building reproducible benchmarks, enforcing physical consistency, interpretability, honest uncertainty estimation, and fair comparisons against well-tuned specialized models. And scale cuts both ways — a model trained at scale can just as easily spread a bias at scale. The broader foundation-model literature keeps making the same point: whatever a downstream system inherits from its foundation includes the flaws and the strengths [10].

Recap: reusable does not mean universal

Two stages carry the whole idea. Pretraining learns a broad waveform representation from the structure in a large archive. Adaptation then teaches that representation a specific scientific task, with local labels and honest evaluation. PhaseNet remains a landmark supervised picker; SeisLM and U-Trans sketch one path toward a representation that many tasks can share.

The real shift is not from “old AI” to “new AI.” It is moving from rebuilding a narrow model for every task toward asking whether one shared representation can make those specialized models easier to build, adapt, and audit. That promise only pays off with the regional validation and uncertainty work that turns a compelling demo into a tool people can trust.

Where to go next

• Try a supervised baseline with the SeisBench documentation and the related phase-picking tutorial.
• Read the companion overview: Modern Seismic Monitoring Systems.
• For a broader survey of the field, start with the earthquake-seismology review in reference [5].

References

1. PhaseNet: A Deep-Neural-Network-Based Seismic Arrival Time Picking Method — Zhu & Beroza (2019), Geophysical Journal International.
2. Earthquake Transformer—an attentive deep-learning model for simultaneous earthquake detection and phase picking — Mousavi et al. (2020), Nature Communications.
3. SeisBench—A Toolbox for Machine Learning in Seismology — Woollam et al. (2022), Seismological Research Letters.
4. Which Picker Fits My Data? A Quantitative Evaluation of Deep Learning Based Seismic Pickers — Münchmeyer et al. (2022), JGR: Solid Earth.
5. Machine Learning in Earthquake Seismology — Mousavi & Beroza (2023), Annual Review of Earth and Planetary Sciences.
6. SeisLM: a Foundation Model for Seismic Waveforms — Liu et al. (2024), arXiv preprint.
7. U-Trans: a foundation model for seismic waveform representation and enhanced downstream earthquake tasks — Saad, Chen & Alkhalifah (2026), Scientific Reports.
8. SeisMoLLM: Advancing Seismic Monitoring via Cross-modal Transfer with Pre-trained Large Language Model — Wang et al. (2025), arXiv preprint.
9. Testing Audio Compression Autoencoders for Seismology: Moving Toward Foundation Models — Laurenti et al. (2026), JGR: Machine Learning and Computation.
10. On the Opportunities and Risks of Foundation Models — Bommasani et al. (2021), arXiv preprint.