Earthquakes and gravity signals
Earthquakes take a heavy toll on human life and property every year. Earthquake early warning systems are therefore essential for high seismic risk zones. However, the existing systems, which are based on seismic and geodetic data have high uncertainty and cannot rapidly estimate the size of an earthquake owing to the slowness of the seismic waves.
To improve earthquake early warning systems, Andrea Licciardi at Université Côte d’Azur, France, and colleagues applied a deep learning method (a subset of machine learning based on artificial neural networks) to prompt elastogravity signals (PEGS). PEGS, which can be detected using seismometers and gravimeters, are transient perturbations to Earth’s gravity field caused by large earthquake ruptures. The signals propagate at the speed of light, much faster than the fastest seismic wave (P-wave), hence solving the problem of delay. The research team trained an algorithm using a database of 350,000 artificial seismograms augmented with empirical noise. Using seismic data from Japan, and specifically the pre-P-wave portion of the seismogram (usually considered as noise) and early portions of the seismogram, the researchers showed that they could rapidly estimate the location and size of earthquakes, both during the rupture (while the earthquake evolves and before the arrival of the complete seismogram) and at the end (the final magnitude, after arrival of most of the seismogram). That is, they were able to track the growth of earthquake rupture, instantaneously and in real-time, as it unfolds. The results show an accuracy of 90% (within 40 seconds from origin time) for earthquakes with magnitudes of MW≥8.6 and 60–70% (within 150 seconds from origin time) for earthquakes with magnitudes of MW 8.2 to 8.6.
The team test the model with a retrospective analysis of real data from the 2011 Tohoku, Japan earthquake. They show that this approach does not suffer from magnitude saturation and can distinguish between large earthquakes (e.g., distinguishing MW 8.0 from MW 9.0). The model out-performs other early warning systems and could be critical for tsunami early warning.
R. Arun Prasath
DETAILS: Nature 606, 319-324 (2022); https://doi.org/10.1038/s41586-022-04672-7