Machine Learning Models for Predicting Ground Motion Amplifications in Japan

Session: Near-Surface Effects: Advances in Site Response Estimation and Its Applications
Type: Oral
Date: 4/28/2020
Time: 05:30 PM
Room: 110 + 140
Description: 

Earthquake-induced ground motions can change their amplitudes, frequency and duration by various factors such as site condition, magnitude and source to site distance. The amplifications of ground motions can significantly affect structures on the ground surface. Therefore, many researchers have developed models to predict ground motion amplification. We used the classical regression method and 3 machine learning techniques (i.e., random forest, neural network and gradient boosting) to develop ground motion amplification models for Japan based on 115,501 ground motions recorded at the KiK-Net station in Japan. We considered various variables such as peak ground accelerations (PGAs), moment magnitude (M), average shear wave velocity in the top 30m (VS30), borehole depth and rupture distance (Rrup). The performances of the models based on the classical regression and machine learning techniques were evaluated for four periods (i.e., T=0.01s, 0.20s, 1.00s, 3.00s) using root mean square errors (RMSEs). The RMSE for the random forest based model was the smallest and that for the classical regression based model was the largest. In addition, VS30 was the most important variable in short period (0.01s-1s), and borehole depth is the most important variable in long period (1s-10s) for predicting ground motion amplifications. The machine learning models based on large amounts of data can provide better prediction for ground motion amplifications than the classical regression models. We generated ground motion amplification maps in Japan based on the random forest based model for two earthquake scenarios (i.e., M5, Rrup = 20km and PGA = 0.05g; M7, Rrup = 10km and PGA = 0.3g). In short periods (T=0.01s and 0.20s), the amplification was generally larger than those in longer periods (T=1.00s and 3.00s).

Presenting Author: Hwanwoo Seo

Authors