Application of Artificial Intelligence PhaseNet and GaMMa, and NonLinLoc Program in Optimization of Regional Tectonic Earthquake Identification on The Island of Java

Abstract The condition and amount of earthquake data in the world continues to increase along with the increasing capacity and coverage of the global seismic network. Therefore, there is a need for technology that can accommodate the increasing amount of data and process it quickly to help improve data processing efficiency in identifying seismic wave phases. This can help the development of more advanced systems, such as disaster early warning systems in Indonesia. In this study, we apply artificial intelligence, namely PhaseNet and GaMMA, along with probabilistic location determination using NonLinLoc, to optimize the identification of regional tectonic earthquakes in Indonesia, particularly on the Island of Java. Seismogram data from 50 regional tectonic earthquakes on Java Island, Indonesia in 2020, with a magnitude range of 3.28 to 5.49, were used. PhaseNet processing of the seismogram data resulted in P-wave and S-wave picks, where 189 out of 569 picks (33%) demonstrated a time difference of under 3 seconds relative to the BMKG catalog, indicating good accuracy. GaMMa generated a total of 44 event associations from a total of 2207 picks out of 50 events (88%) generated by PhaseNet. After filtering based on the Wadati diagram, the arrival times of 20 events were input into NonLinLoc. Of the 20 input events, 17 events had their hypocenters estimated using NonLinLoc. The comparison between NonLinLoc and the BMKG catalog shows that the minimum, average, and maximum differences for determining the hypocenter location respectively are 3.914 km; 206.270 km; and 57.370 km, with 58% of the events having differences in the range of 20-35 km. In summary, using high-quality input data, the implementation of PhaseNet, GaMMa, and NonLinLoc can optimize the identification of regional tectonic earthquakes on Java Island.