Source: Journal of Geophysical Research: Machine Learning and Computation
One seismometer is often not enough to reliably detect earthquakes or human activity such as underground nuclear tests. Rather, researchers combine readings from seismometers distributed across a small geographic area to gain confidence in their analysis. Artificial intelligence (AI) can put together readings from multiple sensors more effectively than classic technology, enabling more reliable detection of weak seismic signals, a new study by Köhler et al. shows.
The researchers leveraged 30 years of readings from seismic arrays operated by the Norwegian research foundation NORSAR and other operators, and they trained an AI model in three different ways to detect seismic signals. First, they trained it on data from one individual station at a time, then applied the model and combined the results from each station. Second, they combined the signals from multiple sensors at the same array using a classic technique, then trained the model on these combined signals from multiple stations. And third, they gave the model all the data from all the array stations and let it decide how to combine them.
The second method (combining signals prior to training) amplified weak signals and provided the most accurate signal detection of all three methods. Meanwhile, the third model (letting the model decide how to combine the station data) was the most computationally efficient strategy, and it fell in between the other two methods in terms of accuracy.
Taking into account the need to balance accuracy with speed, the researchers recommend letting the model decide how to combine data when doing real-time monitoring but combining the data before or after model application when a slower approach is acceptable.
The model does not generalize well to areas outside those it was trained on, however. The reason is that a regionally limited training dataset was used; training on global data is expected to improve results. The problem mainly occurred for S waves, whereas P wave detection generalization was not an issue.
Overall, the results show that AI can improve seismic monitoring by helping researchers detect weak signals from earthquakes, underground nuclear tests, and other seismic activity that might otherwise be difficult to identify. (Journal of Geophysical Research: Machine Learning and Computation, https://doi.org/10.1029/2026JH001249, 2026)
—Saima May Sidik (@saimamay.bsky.social), Science Writer
