Select Page


Indonesian soldiers and search and rescue teams carry the body of a victim

Forecasting aftershocks after an earthquake is difficult. A self-learning algorithm could help here. Artificial intelligence can detect geophysical patterns that our current classical systems cannot see.

Earthquakes are among the most serious natural disasters that can hurt us. They usually come abruptly and develop a great destructive power, as was seen recently in Lombok. The tremors are often accompanied by one or more aftershocks, or even a long series of aftershocks over several days or weeks, which makes the work of rescue workers even more difficult. An aftershock can even do a lot more damage. Therefore, we must try to predict its properties and characteristics as accurately as possible.

Thanks to empirical methods, at least frequency and strength can be predicted quite well. The question of where the earth will tremble again, however, can hardly be answered because the processes in the earth’s crust are too complex for a reliable prediction. Geophysicists at Harvard University in Cambridge have now shown that artificial intelligence has the potential to improve aftershock forecasts.

Phoebe DeVries and her colleagues have developed a self-learning algorithm that identifies recurrent patterns in a large database to establish where aftershocks could occur. The researchers trained their system by feeding it with the geological data from 131,000 events of recent history.

They then tested the predictive ability of their algorithm on 30,000 known examples. The result: The AI ​​was actually able to trace the places where the aftershocks occurred better than the established, purely statistical methods. Nevertheless, the researchers are reluctant. The system is not yet ready for practice. It’s too slow for real-time predictions. In addition, the geophysical data in the databases are still subject to numerous uncertainties.

Another drawback is that the geological processes that lead to aftershocks are currently still being described too simplistically by the theoretical models. The reality is more complex, and at the moment even an earthquake forecast based on AI can only help a little.