2D SH SEISMIC WAVEFIELD ESTIMATION NEAR EARTHQUAKE SOURCE BASED ON PHYSICS-INFORMED NEURAL NETWORKS

Abstract

 It is important for government agencies to quickly assess the damage after an earthquake and to take appropriate emergency actions. Currently, estimates of the spatial distribution of seismic intensity are released immediately after the occurrence of an earthquake, however the key issue is that they underestimate the seismic intensity near the earthquake source fault, which is considered to be larger than that recorded by seismic stations. In this study, to solve the underestimation of seismic intensity near the earthquake source fault, spatial interpolation of ground motion was performed using Physics-Informed Neural Networks (PINNs). The model was trained on a two-dimensional SH wave field using the prediction error of ground motion at assumed station locations, the error obtained from the two-dimensional wave equation, and the stress condition on the ground. The analysis results showed that the model was able to predict values larger than the maximum values recorded by the stations, confirming improved prediction accuracy in the vicinity of the earthquake source. In addition, the method was found to be effective in predicting ground motion before they reach the stations. This method is expected to contribute to the rapid identification of earthquake damage and its response.