American and British geologists have created a new artificial intelligence system capable of predicting earthquakes and have successfully tested it in a laboratory earthquake simulator, according to an article published in the journal GRL.
“For the first time, we've been able to use a machine learning system to analyze acoustic data and predict an earthquake long before it actually happens. This allows us to get enough time to warn and evacuate the population in a timely manner. It's amazing what opportunities artificial intelligence provides us,”said Colin Humphries of the University of Cambridge.
Earthquakes and other dangerous cataclysms associated with the Earth's interior most often occur at the boundaries of faults between tectonic plates, the movement of which is often impeded by irregularities at their edges. When the movement of the plates stops, potential energy accumulates at the point of their contact, which can be released in the form of heat and powerful bursts of acoustic waves at the moment when the rocks in these irregularities cannot withstand and break.
Scientists have long been trying to understand what processes control the accumulation of this energy, and are also looking for ways to "see through" the interior of the Earth so that we can learn about the appearance of such zones of tectonic stress and predict by their properties the probability, strength and timing of new tremors.
Despite enormous progress in this area, such predictions are still extremely inaccurate, which often gives rise to disputes between scientists and politicians who do not like ambiguity. For example, seismologists who incorrectly predicted the magnitude of the earthquake in L'Aquila in Italy in 2009 received real prison sentences for "misinformation" of the population and the death of about three hundred people. This further demotivates seismologists and other scientists to make any specific predictions for the future.
According to Humphreys, one of the reasons why current earthquake predictions are inaccurate or erroneous is that seismographs and other observing devices receive countless signals, only some of which are associated with the accumulation of energy at the boundaries of faults, while others are generated by other phenomena., not connected in any way with tectonic processes.
In some cases, these "obstacles" can be weeded out - and then the forecast is quite accurate, and in other cases, like the disaster of 2009, failure in this regard ends in an unpredictable way.
Similar problems, as Humphries and his colleagues noticed, today are being solved by representatives of a completely different science - computer engineers who develop various systems of machine learning and artificial intelligence. A key feature of modern neural networks is that they can analyze very "dirty" data and find in them what is required to solve a problem: for example, for sorting pictures of cats and dogs or speech recognition in a noisy room.
Promotional video:
Guided by this idea, scientists have created a special "earthquake emulator" at the Los Alamos National Laboratory in the United States, which completely simulates what happens in the faults when new tremors are born, and used it to teach the neural network to "see" the traces of future earthquakes in the dataset that seismographs collect.
After some time, the machine learned to correctly predict "laboratory" earthquakes with a very high degree of accuracy and reliability - this, according to scientists, shows that similar methods can be used to predict the real seismic situation. On the other hand, the current algorithm, most likely, cannot yet be used for these purposes, since it was "trained" not on real data, but on their imitation, and therefore its forecasts can be rather inaccurate when working in the field.