Improving earthquake catalogs helps us better understand how the Earth crust moves and where future earthquakes might happen. In areas where we don’t know much about local faults, having a dense network of seismic stations can help detect small earthquakes that would otherwise be missed.

This study looked at data from the OTRIONS network, active in southern Italy since 2013. Using modern machine learning tools, researchers analyzed data from April 2013 to June 2025. They used a deep learning program called PhaseNet to identify the arrival of seismic waves and another tool, GaMMA, to group those into individual earthquakes. This process found about 27,000 events.

A non-linear algorithm, NonLinLoc, was used to determine where these earthquakes occurred. After manually checking the results, about half of the events were confirmed as real earthquakes. The rest were quarry blasts, errors, or located outside the network.

Compared to a previous catalog created with older methods, this approach found many more events, especially small ones. The results also confirmed known patterns in the Gargano Promontory area, where earthquake depth increases northward and stops at the base of the crust. This work shows how machine learning can greatly improve earthquake monitoring.