AI revolutionises earthquake detection with 10x improvement in sensitivity

TL;DR: Machine learning models have automated earthquake detection over the past seven years, identifying 10 times more earthquakes than traditional methods whilst detecting tremors too small for human analysts to spot in noisy urban environments.

AI tools based on computer imaging have almost completely replaced one of seismology’s fundamental tasks: detecting earthquakes. What previously required teams of undergraduate analysts examining seismograms can now be accomplished automatically by machine learning models like Earthquake Transformer and PhaseNet.

Context and Background

These neural network models, trained on datasets containing over 1.2 million human-labelled seismogram segments, excel at identifying the characteristic “shape” of earthquake waveforms. The Stanford Earthquake Dataset (STEAD), explicitly inspired by ImageNet’s role in launching the deep learning boom, provided the training foundation for current detection systems.

In 2019, researchers at Caltech used template matching to identify 1.6 million earthquakes in Southern California—10 times the previously known total. Modern AI models achieve similar detection rates whilst requiring significantly less computational power, running on consumer CPUs rather than requiring hundreds of GPUs.

Looking Forward

Applications extend beyond simple detection. Researchers used AI-generated earthquake catalogues to create detailed images of Hawaiian volcanic systems, revealing previously hypothesised magma connections between deep sill complexes and shallow volcanic structures. This precision enables better real-time monitoring and more accurate eruption forecasting.

The technology proves particularly valuable for Distributed Acoustic Sensing (DAS), which produces hundreds of gigabytes of data daily from fibre-optic cables. AI tools make processing these massive datasets practical, opening possibilities for identifying entirely new signal types.

However, earthquake prediction—the field’s ultimate goal—remains elusive. Scientists emphasise that whilst AI has revolutionised detection and classification, predicting when the next major quake will strike still proves impossible.

Source Attribution:

Share this article