Currently, predicting precisely where and when an earthquake will take place and what its magnitude will be is not possible. However, in recent decades the study of earthquake precursor phenomena, i.e. the variations in chemical and physical parameters that take place before an earthquake underground, has made much progress. Furthermore, more and more researchers are relying on machine learning (or machine learning), a branch of artificial intelligence, to try to make predictions. The results of the studies carried out so far, even in Italy, are surprising: in some cases it was possible to predict the occurrence of the tremors a few days in advance, even if the epicenter was not precisely identified. This gives hope that in the coming decades, thanks to artificial intelligence, it will be possible to limit the most serious consequences of earthquakes.
How machine learning can help predict earthquakes
The machine learning is a branch of artificial intelligence that uses artificial neural networks, a mathematical model inspired by the networks made up of neurons present in the brains of humans and animals. The purpose of this network is process information And find solutions. To “train” the artificial neural network, so that it improves its ability to carry out processing, they are needed huge amounts of data. In the case of earthquakes, the data provided to the network is collected in the most seismic areas of the planet, where the strongest earthquakes occur, i.e. at the border between two lithospheric plates.
This information, which is associated with the seismograms, includes for example the type of seismic waves recorded, the time they take to reach the detection station, the location of the hypocenter and the magnitude of the earthquake. One of the limits of this system is represented by the fact that seismology is a rather young science and therefore amount of data at our disposal is relatively limited. For example, for the San Andreas Fault in California there are only twenty years of data available that is precise enough to be used in AI prediction.
San Andreas Fault, how it is made and what the Big One is
Where are we with earthquake prediction using artificial intelligence?
Among the studies that have attracted the greatest interest in recent years is that of the geophysicist Paul Johnson, of Los Alamos National Laboratory, New Mexico. Together with his team, the researcher applied machine learning to earthquakes generated in the laboratory: by recording the sounds produced by a rock sample cut by an “artificial fault” and subjected to pressure, the program was able to predict how long it would take for them to occur. check the shocks.
Other researchers fromStanford University they provided the neural network with data about more than 36,000 earthquakes which hit Ridgecrest, California in 2019. In this way they were able to detect seismic tremors a few seconds before they occurred.
In 2023 theUniversity of Texas at Austin, during a seven-month trial in southwest China, trained the AI with data from five years of seismic recordings (including real-time seismic data). The result was that the algorithm has Successfully predicted 14 earthquakes even a week before they occurred, within a radius of approximately 300 km from their actual epicentre, while he predicted 8 of them which did not then occur.
Also in Italy earthquakes are studied with artificial intelligence. In 2022 the National Institute of Oceanography and Experimental Geophysics (OGS) together with the National Institute of Geophysics and Volcanology (INGV) provided a algorithm machine learning (called NESTOR) data from the California seismic catalogs to evaluate the probability that an earthquake with a magnitude greater than 4 is followed by other strong earthquakes. The algorithm correctly predicted, even well in advance, the occurrence of strong tremors.
Machine learning is also very useful in identifying smaller magnitude earthquakes in available seismic data that had not previously been detected by seismologists. In this way existing seismic catalogs can be expanded. An example is the earthquake of 6 April 2009 which hit L'Aquila: the AI managed to identify 114,229 earthquakes with a magnitude of up to 5 during 2009, while the previous catalog contained only 64,000. Having more data available means improving the prediction capabilities of algorithms.
Overall, all these results show that there is a large margin for improvement in earthquake prediction, even if the research is still long.
