News

Faster earthquake early warning with artificial intelligence

Eingestürzte Klosterkirche der Benediktiner in Norcia, Italien. (CCO)
Retrospective evaluation of the method for the 2016 magnitude 6.5 Norcia earthquake in northern Italy. a. 3.5 s after the nucleation of the earthquake, only a few stations in the immediate vi-cinity have recorded the first signals of the earthquake. The outer circle indicates the position of the fast-moving P wave; no signal of the earthquake is yet visible outside this circle. b. The "Transformer Earthquaker Alert Model" receives waveforms and geographical po-sitions of the stations as input and provides probability distributions of ground accelera-tion (shaking) as output. c. For any given point, a probability distribution of the expected ground acceleration is calculated, here visualised for 6 locations (marked in a.) outside the P wavefront. d. Estimation of the 'shake map', which would have been available only 3.5 s after the earthquake (plus a small delay for data transmission) and which produces a specific warning level for each point on the map, which is exceeded with a probability of 40%. (% g corresponds to percent of acceleration due to gravity; significant damage usually occurs from about 10 % g). (Credit: Fig. 3 from: Münchmeyer et. al. The transformer earthquake alerting model. Geophysical Journal International. 2020)

Earthquakes are among the deadliest natural hazards. This is because they release tremendous energies, and because of their unpredictability. There is, however, one property that offers the potential for valuable seconds of warning time: An earthquake sends out waves at different speeds - and the first ones, the so-called P-waves, are far less destructive than the S-waves which arrive later. The further away buildings and infrastructure are from the earthquake's source, the more seconds there are, for example, to close gas lines, cut off the power supply, turn traffic lights to red in front of bridges and stop trains. Traditional methods of early warning, however, are relatively inaccurate or offer only very short warning times.

As part of an interdisciplinary collaboration between the German Research Centre for Geosciences in Potsdam and the Humboldt-Universität zu Berlin, Jannes Münchmeyer and his colleagues adapted an artificial intelligence method originally from the field of text comprehension and automated translation, so-called transformer networks, to the analysis of seismic data. Their goal was to achieve faster and often significantly more accurate predictions of the shaking to be expected in the vicinity of an earthquake.

The new method was tested with data sets from the highly earthquake-prone countries of Italy and Japan, both of which have a very dense network of earthquake stations. As is usual with "machine learning", the method was tested on a set of thousands of recorded quakes and, in a sense, adjusted. Afterwards, the researchers provided another set of recorded quakes that the algorithm did not "know" yet. In these retrospective tests with actual earthquakes, there was a significant improvement in warning accuracy. Jannes Münchmeyer says: "We would have gotten a higher number of correct and fast warnings of stronger tremors than with previous approaches."

Another advantage of the new method, he says, is that an estimate of the accuracy of the prediction is also calculated. Frederik Tilmann from the GFZ explains: "This means that thresholds at which warnings are given can be adapted to local conditions and people's needs."

The interdisciplinary work of seismologists and data scientists (computer scientists and mathematicians) was crucial to the success of this research approach and was made possible by the „Helmholtz Einstein International Berlin Research School in Data Science” (HEIBRiDS) initiated by the Helmholtz Association and the Einstein Centre for the Digital Future, as well as the Geo.X-Network.

Original study: Jannes Münchmeyer; Dino Bindi; Ulf Leser; Frederik Tilmann, 2021. The transformer earthquake alerting model: A new versatile approach to earthquake early warning. Geophysical Journal International. DOI: 10.1093/gji/ggaa609

Scientific contact:

Jannes Münchmeyer 
Section Seismology
Helmholtz Centre Potsdam GFZ
German Research Centre for Geosciences
Telegrafenberg
14473 Potsdam
Phone: +49 331 288-1228
Email: jannes.muenchmeyer@gfz-potsdam.de

Prof. Frederik Tilmann  
Section head Seismology
Helmholtz Centre Potsdam
GFZ German Research Centre for Geosciences
Telegrafenberg
14473 Potsdam
Phone: +49 331 288-1240
Email: frederik.tilmann@gfz-potsdam.de

Media contact:

Josef Zens  
Head of Public and Media Relations
Helmholtz Centre Potsdam
GFZ German Research Centre for Geosciences
Telegrafenberg
14473 Potsdam
Phone: +49 331 288-1040
Email: josef.zens@gfz-potsdam.de

Additional News

Diamonds need Voltage

Fragen und Antworten zum Erdbeben in Kroatien

Melody of an Alpine summit falling apart

Geothermal exploration with fibre optic seismics in Potsdam

Early warning system for dams at risk

Better calculation of stress in the underground

How hot can it be for life to exist deep below the ocean floor?

Earthquake Scenario for Large German City

Potsdam Young Scientist Award for David Uhlig

VARDA: New database for looking into Earth's past climates, one layer at a time