PREDICTION ON THE AMPLITUDES OF SEISMIC UNDERGROUND MOTIONS BASED ON DEEP NEURAL NETWORK

The underground structures were significantly damaged in the 1995 Kobe Earthquake, and since then the seismic safety of underground structures has received lots of attentions. In the dynamic nonlinear analysis of underground structures, one of key issues is how to determine the amplitudes of seismic underground motions (e.g., peak acceleration of underground record). Because the number of recorded underground records is far less than that of ground motions, there is no reliable prediction tool for the amplitude of seismic underground motions. Recently, the deep learning has become a popular and powerful tool to predict the behavior of complicated systems. The amplitudes of seismic underground motions depend on the characteristics of soil and earthquake, making the prediction very complicated. This paper aims to predict the peak acceleration of seismic underground records using the deep neural network (DNN). The total of 86880 underground records (including two horizontal components for each station) recorded by 4258 earthquake events (magnitude varies from 4.0 to 9.0) on 639 stations (epicentral distance varies from 0.74 km to 135 km), are collected from the Kik-net database. The average shear wave velocity of top 30 m soil (V_S30) varies from 111.11 m/s to 2100 m/s, and the peak ground acceleration (PGA) varies from 1.57 gal to 1080 gal, in order to cover the nonlinear site response. The seismic underground records in the east-west (EW) direction (i.e., the total of 43440 records) are randomly split into training (80% of the database), validation (10% of the database) and test (10% of the database) datasets, respectively, to check the overfitting of the DNN model (consisting of 5 layers and 256 neural in each layer) by monitoring the loss of the validation and testing datasets. The Adam optimizer is used as the optimization algorithm to reduce the error of the output with a batch size of 512 and 150 epochs of training. Note that the deep learning in this study was performed using Tensorflow. The input parameters include magnitude, epicentral distance, depth of the underground receiver (varies from 99 m to 2003 m), the classic site-characterization term Vs30, PGA, peak ground velocity (PGV) and peak ground displacement (PGD). The generalization ability of DNN model based on the database in the EW direction is also studied with the database in the north-south (NS) direction, and the results are summarized in Table 1. The results in Table 1 indicate that the DNN model provides good prediction on the peak acceleration of seismic underground records, and the generalization ability of DNN model is also accepted.