What if artificial intelligence (AI) could predict the date, time and magnitude of future earthquakes? A team of French researchers from the École normale supérieure (ENS) – PSL is developing ” AI algorithms that continuously scan data from seismometers recording ground vibrations, in an attempt to identify specific behaviors that precede earthquakes “. This Wednesday, August 19, this research is the subject of a publication in the scientific journal Nature Communications. The results are as encouraging as they are unprecedented.
Record “recurring behaviors”
” The concept is simple: by analyzing the seismic records preceding many earthquakes, an algorithm can identify recurring behaviors in these recordings, and learn to recognize the next occurrence of the rupture », Write the two doctoral students Claudia Hulbert and Romain Jolivet in a popular article published on The Conversation.
In short, they use the Machine learning in an attempt to understand “seismic cycles” – defined as ” the alternation of earthquakes and periods during which tectonic stresses accumulate “. Determine a pre-earthquake pattern to better prepare for it.
If the concept is simple at first glance, the method is sophisticated. ” This approach was Originally developed on “laboratory earthquakes”: by compressing samples of steel or rock in a press and sliding them against each other, we can give rise to thousands of fractures, each analogous to an earthquake, the only difference being the size “, Explain the researchers. ” For example, a magnitude 7 earthquake corresponds approximately to a rupture of about 100 km long for a few tens of centimeters of slip, while laboratory earthquakes are only a few centimeters long for a few millimeters of slip. “. Their approach is therefore based on scaling their observations in the lab.
The algorithm is then ” trained “By” vibrations emitted by the sample before the ruptures “Observed in the laboratory for” identify whether there are characteristic signals of an upcoming rupture and, above all, recognize them “… full-scale.
Testing on Vancouver Island
Nevertheless, researchers remain cautious. Their ” method is not yet applicable to large, devastating but rare earthquakes that make the headlines “, They warn. Therefore, they experimented with their earthquake-detecting AI on Vancouver Island, Canada, known for its ” slow earthquakes “. These types of earthquakes can last for weeks or months and their effects are not disastrous.
However, these ” slow earthquakes Have some commonalities with classic earthquakes – for example they can be this big and the warning signs are the same. The success of the algorithm developed by French researchers for these phenomena therefore opens up a field of possibilities: ” we might be tempted to simply scale it up: a slow earthquake lasts a month and we see warning signs 100 days (or three months) before the earthquake. A classic high magnitude earthquake lasts 10 to 100 seconds and these precursor signals would therefore be observable 30 to 300 seconds before… “, suggest the researchers before asking:” Is this the case? Would it help to have an estimate at such short notice? “. Research should intensify.