Combining physical model with neural networks for earthquake site response prediction

Abstract

The amplitude of seismic waves will be significantly amplified near the Earth's surface, and this phenomenon is known as the seismic site response. Site response prediction is of paramount importance for the seismic-resistant building design and seismic risk assessment. However, accurately predicting site response has always been a challenge due to the incomplete physical knowledge and insufficient dataset volumes. Here, we propose an approach that combines the neural networks with classical homogeneous layered model for site response prediction. This approach exploits the potential for improving the accuracy of site response prediction from both the physical and data perspectives, which reduces the requirements for the model complexity and the training data volume. Compared to the physics-driven method, this approach reduces the estimation errors by approximately 50 % on average, and corrects the correlation between the observed and predicted results. This approach firstly reproduces the four-stage characteristics of the site response in the entire seismic band, and provides a new framework for site response prediction.