Development of far-source earthquake ground motion model using recurrent-based neural network

Countries having low to medium seismicity experience rare large, far-distance earthquakes. The development of its ground motion model is difficult due to data scarcity. Thus, the applicability of the LSTM recurrent neural network is explored, due to its ability to predict sequential data. Earthquakes from databases are collected with magnitudes larger than 6.5 Mw and distances between 200 and 500 km. The network is constructed with five input data, two LSTM layers, and 23 sequential outputs. The results show that the network underestimates the amplification at lower periods while predicting larger motions at higher periods, typical for far-source earthquakes.

Actions (login required)

View Item