A new journal publication in Nature Communications is shaking up the science of predicting earthquakes. By using automatic speech recognition designed to encode waveforms for translation, researchers at Los Alamos National Laboratory were able to modify this voice-to-text AI to correctly predict the timing of a slip during a repeating collapse sequence producing approximately magnitude-5 earthquakes at the Kīlauea volcano on Hawai’i. The creative approach to applying this AI technology may well be a step toward a more efficient earthquake monitoring system.
“It is similar to technology used in speech-to-text that is included in many apps,” said Christopher Johnson, a research scientist at Los Alamos. “Instead of translating audio recordings to words, we map the input seismic waveforms to a deep learning model trained specifically for the task of predicting the timing of slip.”
Johnson built upon Wav2Vec-2.0, an automatic speech-recognition AI from Facebook AI Research, to power this research.
The modeling approach performs surprisingly well at predicting real-time slip events but underperforms when predicting future slip events. Regardless, the real-time prediction results are a significant advancement. Further research will work toward increasing the future time frame in which the technique is accurate.
While voice-to-text software may seem like an unusual choice for seismic analysis, the waveforms actually make it a natural fit.
“Audio data for automatic speech recognition are analogous to continuous seismic waveforms,” Johnson said. “The concept of encoding the signal to a high-dimensional representation, then applying a transformer network to make predictions, is a continuation of work we developed using laboratory earthquake experiments.”
As scientists continue these efforts to better predict seismologic hazards, voice-to-text AI could be an integral and surprising solution for more accurate predictions. Research in this domain to improve earthquake monitoring capabilities will help keep communities across the country safer.
Paper: “Automatic Speech Recognition Predicts Contemporaneous Earthquake Fault Displacement.” Nature.
Funding: U.S. Department of Energy, Office of Science, Office of Basic Energy Sciences
LA-UR-25-21103
