Deep Learning for Site Response Estimation from Geotechnical Array Data

Session: Near-Surface Effects: Advances in Site Response Estimation and Its Applications
Type: Oral
Date: 4/28/2020
Time: 11:15 AM
Room: 110 + 140
Description: 

Ground motions recorded on vertical arrays show that theoretical methods for site response prediction frequently fail to reproduce the observed surface-to-borehole amplification function due to modeling simplifications. Although sophisticated 3D simulations with realistic representations of laterally heterogeneous media may result in improved predictions through individual case studies, these attempts typically require computationally expensive trial-and-error parameter variations. Instead, we propose to improve site response prediction by harnessing the large amount of strong motion data collected on geotechnical (borehole) arrays as a training dataset for a deep learning algorithm. The goal of our project is to derive an independent method for the assessment of site response that is not relying on simplifying assumptions or proxies. Our approach is based on an artificial neural network (ANN) consisting of multiple layers, where the input layer contains the discretized shear-wave velocity profile and frequency of amplification. The output layer consists of  a single neuron representing the amplification at the selected frequency. Experiments with 600 real KiK-net soil profiles show that a properly regularized ANN with seven hidden layers is able to learn the theoretical SH amplification functions and predict the theoretical site response for profiles that were not part of the training set. We train the network using a large number of observed amplifications collected at vertical arrays located in California and Japan. We also feed the network with other parameters which reflect the characteristics of the incoming wavefield and affect the linear or nonlinear response of the site, such as earthquake magnitude, rupture distance and shaking duration. We experiment with different network architectures and regularization strategies to improve the predictions and to avoid overfitting. The accuracy of the network is evaluated against traditional (theoretical) site response assessment techniques for randomly chosen sites and earthquakes which were not part of the training dataset.

Presenting Author: Daniel Roten

Authors