Large earthquakes: can we detect them more quickly using AI?


The provisional toll from the violent earthquake which struck Morocco during the night from Friday to Saturday continues to increase. There are now more than 2,000 dead, the Interior Ministry said.

According to the National Center for Scientific and Technical Research, the epicenter of the earthquake was located in the province of al-Haouz, southwest of Marrakech and the earthquake caused significant damage in several cities

Over the past thirty years, earthquakes and the tsunamis they generate have caused the deaths of nearly a million people . If prediction as such of these events is impossible , warning systems have been put in place to limit the human and material cost of these disasters.

These systems do not predict the future, but try to detect earthquakes and estimate their magnitude as quickly as possible. Currently, they use seismic waves to try to warn populations a few seconds before the tremors, even if unfortunately the result is not always there.

Tsunamis propagate more slowly, leaving more time to act, which is generally ( a few tens of minutes ). However, warning systems have great difficulty quickly assessing the magnitude of very large earthquakes. For example, the Japanese system estimated a magnitude of 8 instead of 9 during the 2011 earthquake, and therefore a wave of 3 meters instead of 15, an error with dramatic consequences in Fukushima.

In order to improve seismic and tsunami warning systems, we are currently working on an artificial intelligence (AI) algorithm, based on waves of gravitational origin, which estimates the magnitude of large earthquakes more reliably and quickly. .

Seismic warning systems

The seismic signals recorded earliest on seismometers are compression waves (called P waves). These waves propagate at around 6.5 km per second. If you are 65 km further from the epicenter than the nearest sensors, you will therefore feel the first tremors 10 seconds after these sensors have recorded the first P waves. In practice, taking into account the transmission time and processing these waves, your 10 seconds will probably be reduced to 5 or 6.

But the most destructive waves, shear waves (called S waves), propagate slightly more slowly than P waves (at around 3.5 km per second), it is possible to anticipate the strongest tremors by a few seconds. . On this principle, in countries with seismic alert systems such as Japan, when an earthquake is detected near your position, you receive an SMS alert informing you of the imminence of tremors.

Tsunami warning systems

Unfortunately, for both instrumental and fundamental reasons, P waves do not provide reliable information on the magnitude of very large earthquakes. Seismic warning systems, based on these waves, thus prove incapable of making the difference between an earthquake of magnitude 8 and an earthquake of magnitude 9, posing a major problem for the estimation of the tsunami, as illustrated the example of Fukushima in 2011. In fact, a magnitude 9 earthquake is 30 times “bigger” than a magnitude 8 earthquake, the tsunami it generates is therefore considerably larger.

To more reliably estimate the magnitude of large earthquakes, warning systems based on another type of wave, called W phase, have been developed . The W phase has much better sensitivity to magnitude than P waves, but propagates much more slowly. It is recorded between 10 and 30 minutes after the origin of the earthquake, i.e. shortly before the arrival of the tsunami.

The discovery of gravitational signals

In 2017, previously unknown signals were discovered . These signals, called PEGS for “prompt elasto-gravity signals”, have provided a glimpse of a new possibility of estimating the magnitude of large earthquakes more quickly and more reliably.. When an earthquake occurs, a huge mass of rock is suddenly set into motion. This mass of moving rock causes a disturbance in the Earth’s gravity field (gravity). This disturbance is extremely weak, but propagates like a gravitational wave, at the speed of light. Instantaneously on the scale of the Earth. Gravity being an acceleration and seismometers recording the acceleration of the ground, PEGS are recorded by our “classic” measuring instruments. In addition, these signals are very sensitive to magnitude, much more than P waves in the case of large events.

PEGS therefore have the ideal characteristics to power an alert system. However, their detection is made difficult by their very low amplitude. (about a million times weaker than P waves). How can we use such weak signals to alert?

An AI to exploit gravitational signals

Emerging AI technology is proving highly effective at quickly extracting weak signals from large volumes of noisy data. We have developed an AI algorithm which estimates the magnitude of the current earthquake every second from the PEGS] It is therefore necessary to develop new, more reliable and rapid systems, in order to have an update strategy. the most effective shelter possible. We have developed an artificial intelligence (AI) algorithm, based on waves of gravitational origin propagating at the speed of light, to estimate the magnitude of large earthquakes more quickly and reliably., published very recently in Nature. As large earthquakes are rare, we simulated hundreds of thousands of possible earthquake scenarios along Japan’s major faults.

In each scenario, we calculated the expected PEGS on all seismometers in the region and trained the AI ​​to “find” the magnitude and location of the earthquake by giving it the answer each time. We then tested the performance of the AI ​​on the data recorded during the Fukushima earthquake. The results indicate that we could have estimated the magnitude of the earthquake as soon as the rupture ended (i.e. 2 minutes after the origin of the event), and therefore very quickly obtained a much better estimate of the height of the wave.

The results being encouraging, we are now moving on to the phase of implementing the algorithm in an operational alert system, starting with Peru where we are expecting a very big event (which could occur tomorrow as in 300 or 600 years). We are also working to improve the algorithm’s performance for earthquakes of more moderate magnitude. It works in its current version for earthquakes of magnitude greater than 8.3, which already makes it very useful for estimating tsunamis (which only concern these very large earthquakes) but greatly limits the possibilities for warning about tremors. (because the latter are felt in most cases before the earthquake reaches such magnitude).

Finally, we aim to develop a global version of this algorithm which would use seismometers to alert on earthquakes occurring anywhere on Earth, thus offering a particularly interesting global alert “coverage” for poorly equipped regions.

Author Bio: Quentin Bletery is a Geophysicist, IRD Research Manager at the Géoazur laboratory at the Research Institute for Development (IRD)