AI Predicts Earthquakes Before They Strike

For the majority of human history, the typical response to the question “can we predict earthquakes” has been a quiet, somewhat ashamed “no.” Folklore about restless dogs and headaches prior to tremors has been refuted by seismologists for decades. To provide a helpful warning, the crust is too deep, too disorganized, and too quiet. For a very long time, that was the consensus. It’s interesting to note that the consensus is beginning to falter, not because the issue has been fully resolved, but rather because the type of instrument used to listen to the Earth has evolved.

Earlier this year, a team at Kyoto University reported that machine learning could identify the weak signals that appear right before a meter-scale lab earthquake. This type of controlled rock-fracture experiment is used by researchers to simulate natural faults. The study’s first author, Reiju Norisugi, talked about using acoustic emission data—tiny, nearly undetectable foreshocks—to train models and discover patterns that the human eye would never have noticed. The results, which were published in Nature Communications, do not assert that the puzzle has been solved. They imply that something can now read the precursors and that they are real.

Researchers at Los Alamos National Laboratory in New Mexico, on the other side of the Pacific, have been engaging in an odd and possibly more ingenious activity. Seismic data was fed into Meta’s Wav2Vec-2.0 AI model, which was initially designed to transcribe human speech. When you sit with the logic, it makes a sort of poetic sense. Speech is nothing more than temporal patterns. It is also a flaw. The continuous waveform data from the 2018 Kīlauea collapse in Hawaii, when the magma chamber gave way and caused months of minor earthquakes, was fed into the model by Christopher Johnson and his colleagues. The model began predicting the timing of slip events with unexpected accuracy after being trained without manually labeled examples. It performed better than the conventional gradient-boosted-tree techniques, which have long been the mainstays of this type of work.

This is a humorous irony. The machine learning systems that process earthquake data were not designed to do so. They originated from unglamorous consumer technology that people hardly notice, such as voice assistants and transcription services. However, identifying structure in continuous, noisy signals turns out to be the same skill. It’s difficult to ignore how frequently advances in earth science now come from completely different directions.

The British Geological Survey’s partnership with Edinburgh and Padua, which was published in Earth, Planets and Space last November, may be the most operationally useful work. In contrast to the hours or days that the typical ETAS model takes to run its simulations, their AI tool predicts aftershocks within seconds of a major earthquake. That gap is crucial for emergency services. During the first hectic hour, decisions are made regarding building evacuations, hospital reroutes, and crew assignments. The PhD student in charge of the project, Foteini Dervisi, took care to present it as an enhancement rather than a substitute. The speed increase is still difficult to ignore.

Naturally, it’s still unclear if any of this will be able to accurately predict large earthquakes weeks in advance. Faults in the lab are not actual faults. A subduction zone is not the same as a volcanic collapse. In seismology, there is a long, humble list of things that appeared promising but later fell apart when examined. Observing this, however, gives the impression that the question has changed. The question of whether earthquakes can ever be predicted is not being discussed for the first time in a long time. It concerns how quickly, how accurately, and how the world will use that information once it becomes available.