Deep learning prediction model for ground motion amplification effect of sedimentary valleys with varying shear wave velocities

The site effects of sedimentary valleys caused by earthquakes were widely investigated using numerical methods, which had the challenge of the high computational costs. This study explores the feasibility of deep learning methods for obtaining the nonlinear seismic response of sedimentary valleys with varying shear wave velocities. The proposed deep learning model was constructed based on the Long Short-Term Memory (LSTM) network with the hybrid input features, including time series of input waves and shear wave velocities of valleys. The output features of this model were the time series of the seismic response at representative surface locations of sedimentary valleys, and the surface locations were determined via the Principal Component Analysis (PCA). Based on the model validation, the site effect and structural fragility solved by both the proposed LSTM model and traditional one-dimensional (1-D) soil analysis method were compared and discussed. The results indicate that the LSTM model exhibits great efficiency and precision (Coefficient of determination R ²  = 0.96) in assessing the site response and fragility analysis. Compared to LSTM models, the 1-D soil analysis method exhibits a notable underestimation of the structural fragility, with a maximum difference of approximately 50%.