Scientists have developed a system that can estimate the magnitude and location of a major earthquake almost immediately after it occurs by analyzing an extremely weak change in Earth’s gravitational field. Unlike traditional methods, the technology uses a signal that appears before seismic waves, while artificial intelligence helps extract the information needed to assess the scale of the event.

The development has already moved beyond a laboratory experiment: researchers have created an application that allows the algorithm to be integrated directly into the SeisComP seismic monitoring system. The technology is being tested and implemented in different countries, including tsunami early warning systems in Peru and Alaska.

Quentin Bletery, a researcher at the French National Research Institute for Sustainable Development (IRD) and the Geoazur laboratory, presented the operating principle of the system and the results of its tests.

The system is intended to solve the problem of the first minutes after an earthquake

According to him, in the event of a strong undersea earthquake, specialists have a relatively large amount of time before a tsunami arrives — depending on the distance to the coast, the wave may reach it in one to two hours. However, the decision on whether a warning is necessary must be made much earlier. The key parameter here is the earthquake’s magnitude. The potential size of the tsunami largely depends on the scale of the event, the specialist noted. 

The problem is that traditional early warning systems assess an earthquake based on the first seismic waves. For ordinary events, this approach works well, but for extremely large earthquakes a saturation effect occurs: the initial data can produce a significantly underestimated magnitude.

Bletery cited as an example the magnitude 9 earthquake that struck Japan in 2011. Such an earthquake does not develop instantaneously — its full magnitude is reached over roughly two minutes. At the same time, Japan’s early warning system initially estimated the scale of the event as much lower than it actually was and effectively stopped at about magnitude 8.

For tsunami forecasting, this is a fundamental difference. According to the researcher, in this case there may be about a 30-fold difference in tsunami amplitude between magnitude 8 and magnitude 9 earthquakes.

That is why, for early warning systems, it is especially important to obtain information about the true scale of the largest earthquakes as early as possible.

Gravity provides information earlier than seismic waves

After the 2011 Tohoku earthquake, researchers turned their attention to the so-called prompt elasto-gravity signal, or PEGS. When a giant earthquake occurs, enormous masses of rock shift over considerable distances. This causes an extremely small change in Earth’s gravitational field.

The signal is so weak that its magnitude is on the order of 1 nanometer per square second. However, it can be recorded by seismological instruments, because seismographs measure acceleration, and gravitational influence also manifests itself as acceleration. The main advantage of PEGS is its propagation speed. The gravitational disturbance propagates at the speed of light, whereas seismic waves travel through Earth much more slowly.

This creates a kind of time window: information about the change in the gravitational field can be obtained before the arrival of seismic waves. At the same time, the signal contains information not only about the fact that an earthquake has occurred, but also about its scale. Bletery showed that the calculated signals for magnitude 8.5 and 9 earthquakes differ noticeably. In theory, this makes it possible to determine how large the event is even before traditional seismic information provides a complete picture.

Thousands of virtual earthquakes were created for the AI

The main technical problem was that the gravitational signal itself is extremely weak. To teach the algorithm to recognize it, researchers created a large number of virtual earthquakes at various points in Japan’s subduction zones. For each scenario, they calculated synthetic signals that the seismic network would have recorded if such an earthquake had actually occurred. They then added real seismic noise collected over a year of observations to these data. In this way, the scientists obtained an artificial dataset that was as close as possible to real monitoring conditions.

A neural network was trained on these data to determine the magnitude and location of an earthquake from the characteristic distribution of the signal across the seismic network. Later, the researchers moved from a convolutional neural network to a graph neural network (GNN). In this case, the algorithm takes into account not only the signals themselves, but also the geometric arrangement of the seismic stations relative to one another.

This proved especially important for smaller earthquakes, where the gravitational signal is even weaker.

The technology is already being tested in different countries

The method was tested in autonomous mode using data from Japan, Chile, and Alaska. The algorithm is now also being introduced on a trial basis into Peru’s early warning system, is being used in work on a system in Alaska, and has begun to be applied in New Caledonia.

In Alaska, the system was additionally trained to determine the earthquake moment tensor — a parameter that makes it possible to establish the fault geometry and the nature of crustal movement. This is directly relevant for tsunami assessment. Two earthquakes of the same magnitude will not necessarily generate tsunamis of the same scale: much depends on how the displacement of Earth’s crust occurred.

In experiments, the algorithm performed well for earthquakes of about magnitude 8 and higher. For weaker events, the gravitational signal became too small. However, the use of a graph neural network made it possible to improve the stability of the results and expand the system’s operating range.

The algorithm can be integrated into an existing monitoring system

The researchers’ next step is to make the technology part of real seismological monitoring infrastructure. According to Bletery, an application has been created for this purpose that allows the algorithm to be connected directly to the SeisComP platform used by specialists.

Thus, this is no longer just a neural network operating in a research environment. The developers are creating a tool that can be integrated into existing monitoring centers and used alongside traditional methods for determining earthquake parameters.

The system’s task is to obtain an additional estimate of the magnitude and location of a major earthquake almost immediately after it occurs and, going forward, to use this information when deciding whether a tsunami warning is necessary.

At the same time, the new technology is not intended to replace traditional seismic systems. Its purpose is to give specialists an additional source of information precisely at the moment when conventional methods still cannot show the true scale of an extremely large earthquake.

If further tests confirm the results, the gravitational signal combined with AI could become an additional layer of tsunami early warning.