Introduction: a quiet revolution in seismology
On January 1, 2008, Calipatria, California, recorded a magnitude -0.53 earthquake that most people would never feel. Yet, this tiny event is emblematic of a broader shift: over the past seven years, AI-powered image-like analysis has automated one of seismology’s core tasks—detecting earthquakes. These machine learning tools can identify tremors in noisy urban environments and build richer, more accurate catalogs that improve our understanding of the Earth’s interior and potential hazards.
From templates to transformers: two eras of detection
Earlier approaches relied on template matching, where a computer searched for waveform snippets that resembled known earthquakes. This worked well in regions with extensive historical data but was expensive and struggled in data-sparse areas. In 2019, Caltech researchers using template matching uncovered about 1.6 million earthquakes in Southern California, most of them very small, but the method demanded enormous computational power and pre-existing templates.
Enter the Earthquake Transformer and other AI models. These detection systems use neural networks to learn from vast labeled datasets, enabling faster, more generalizable earthquake detection and phase picking. Unlike template matching, AI models can infer new earthquakes in new regions and balance sensitivity with robustness against noise—crucial for urban seismology where traffic, construction, and other sounds complicate signals.
How AI detects quakes and times wave arrivals
Seismographs record three-dimensional ground motion. To determine when an earthquake starts and when the P and S waves arrive, AI models process time series data to output three probabilities at each moment: the likelihood of an earthquake, the arrival of the P wave, and the arrival of the S wave. Modern models, such as Earthquake Transformer, adapt image-recognition techniques to one-dimensional time series, using 0.1‑second input chunks and deeper layers to capture patterns across longer timescales.
The attention mechanism—a staple of large language models—helps the network correlate early P-wave cues with later signal structure, refining both detection and phase picking. This yields a more precise seismic timeline, which in turn improves the interpretation of the Earth’s interior and the behavior of tectonic plates.
The data and the architecture behind the success
Critical to AI success in seismology is data. Large, labeled seismic datasets—such as the STEAD catalog with over a million annotated segments—provide the training ground for models. While architectures vary, many rely on a mix of one-dimensional convolutions and, in some cases, attention layers, to extract meaningful features from noisy waveforms. Some models, like PhaseNet, focus on the arrival times of P and S waves using simpler architectures, while Earthquake Transformer employs a more holistic approach to event detection and phase classification.
These methods aren’t intended to replace seismologists but to augment their work. They produce larger, more complete earthquake catalogs, especially for small events, enabling researchers to study the Earth’s subsurface more comprehensively and to forecast potential hazards with greater confidence.
Beyond detection: what AI changes in seismology
AI-driven catalogs have already transformed volcano monitoring by revealing intricate magma pathways through dense quakes, as seen in Hawaiian seismic studies. They also improve data handling for Distributed Acoustic Sensing (DAS), where fiber-optic cables generate enormous data streams. AI-based phase picking in DAS data has made it feasible to time earthquakes with unprecedented precision, unlocking new downstream analyses and potentially revealing previously unseen signal types.
What remains uncertain—and why the revolution is ongoing
Despite breakthroughs, earthquake prediction remains elusive. AI has largely expanded cataloging and analysis capabilities rather than delivering a reliable forecast. Yet by transforming how we detect and characterize quakes, AI has elevated the realism and utility of seismic data, making the field more capable of responding to future hazards and scientific questions.
Conclusion: a clearer view of the Earth
As seismology continues to adopt and refine AI tools, the discipline moves closer to the aspirational moment: a robust, global, high-resolution understanding of earthquakes and the underlying structure of our planet. It’s like putting on glasses for the first time: the world comes into sharper focus, one tiny tremor at a time.
