Data Study Group Final Report: University of Leeds

Detecting the occurrence and location of earthquakes is vital in a variety of settings. An increase in seismicity (waves generated by earthquakes or rupturing) in volcanoes indicates changes in behaviour which then inform decisions on eruption likelihood and whether evacuation is necessary. Recently, seismic activity has been linked to potential geothermal activity and thus is a promising method for evaluating a location for a geothermal power plant. Ruptures in glaciers generate seismic activity and analysing and locating these signals give a non-invasive method to understanding the dynamics and structure of glaciers. The tragic incidents of the 2023 earthquake in Turkey also motivate the rapid and accurate detection and location of seismicity both before and after the main earthquake; the aftershock earthquakes map out the fault structure beneath the surface and highlight regions where events are likely to take place in the near future.

Automating the detection of seismic arrivals is useful in a variety of fields but it has yet to be perfected and many approaches, including those using deep learning, have been trialled. Techniques that will rapidly detect and locate seismicity automatically are still computationally expensive, require human intervention, or rely on uncertain information about the geological structure. Other studies have used neural networks to extract the arrival time of the seismic waves caused by the earthquakes and then give these times to techniques typically to locate where the rupture occurred. These approaches only use one seismogram at a time and information about the seismometer’s location is rarely used. Furthermore, issues remain with assumptions of accurately predicting the time taken for waves to pass through the subsurface from the earthquake to the seismometer.

This challenge involves both identifying seismic arrivals in time series data of ground motion at seismometers and then locating said arrivals in terms of their latitude, longitude, and depth. Meeting this challenge would mark a huge improvement in our ability to monitor earthquakes and manage the risks they pose. Aspects of this work, such as identifying useful features or the neural network architecture, could be translated to and improve other 2 fields.