Deep Learning-Based Site Amplification Models For Central And Eastern North America

This paper presents deep learning-based site amplification models developed from large-scale simulated site amplification in Central and Eastern North America (CENA). The error evaluation of conventional simulation-based linear and nonlinear response spectrum (RS) and smoothed Fourier amplitude spectrum (FAS) amplification models highlights that fitting whole dataset to predetermined functional forms cannot capture the complex behavior inherent in the simulated amplification in CENA. Deep learning through Artificial Neural Network (ANN) is adopted for a new set of RS and FAS amplification models without the limitations of conventional regression models. This study shows significant improvements over conventional functions by use of ANN-based models: (i) the error in estimation is reduced up to 30% relative to conventional linear and total RS models, (ii) the simulated shallow site response is captured more accurately, and (iii) a continuous model for linear FAS amplification, previously provided as tabulated functions of V-S30 and soil depth, is produced.