Machine Learning-based Models to Predict Ground Motion Intensity in South Korea
South Korea was recently struck by two earthquakes with local magnitudes (ML) of 5.8 and 5.4, which were recorded as the first and second-largest earthquakes in the Korean Peninsula. These two earthquakes are attracting attention for the necessity of seismic hazard evaluations in Korea. Some ground motion prediction models (GMPMs) have been developed in South Korea. Emolo et al. (2015) used 11,129 ground motions recorded during 222 earthquakes between 2007 and 2012 to make the ground motion prediction equation (GMPE). Jeong and Lee (2018) made a GMPE using 115,000 synthetic ground motions generated by the ground motion model developed on the seismic records from 11 earthquakes between 2003 and 2016.
In this study, we developed GMPMs to estimate 5% damped pseudo-spectral accelerations (PSAs) at 27 periods ranging from 0.01 to 10 s based on three machine learning techniques: Random Forest (RF) and Gradient Boosting (GB) with 500 trees, and Artificial Neural Network (ANN) that consists of two hidden layers with 31 and 51 nodes. We used 1,189 ground motions recorded at 50 surface stations in South Korea. We considered five independent variables: ML, epicentral distance (Repi), average shear wave velocity of the upper 30 m (VS30), focal depth and slope angle. The prediction performances of three machine learning-based models were compared with a classical regression-based model (Emolo et al., 2015). Among the four models, the GB-based model showed the best performance with the smallest errors on the unseen data. To compute variable importance, we used Mean Decrease in Impurity function. It turned out that the Repi and ML were the most influential for periods ranging from 0.01 to 0.1 s and periods longer than 0.2 s, respectively. We also applied the GB-based model to recent earthquakes and the estimated PSAs were in good agreement with the records. To enable to implement in seismic hazard assessments, we developed an executable version of the trained GB-based model.
Session: Site Response Characterization in Seismic Hazard Analysis [Poster]
Type: Poster
Room: Evergreen Ballroom
Date: 4/22/2022
Presentation Time: 08:00 AM Pacific
Presenting Author: Jisong Kim
Student Presenter: Yes
Additional Authors
Jisong Kim Presenting Author jisong@unist.ac.kr Ulsan National Institute of Science and Technology |
Hwanwoo Seo hwanwooseo@unist.ac.kr Ulsan National Institute of Science and Technology |
Byungmin Kim Corresponding Author byungmin.kim@unist.ac.kr Ulsan National Institute of Science and Technology |
|
|
|
|
|
|
Machine Learning-based Models to Predict Ground Motion Intensity in South Korea
Category
Site Response Characterization in Seismic Hazard Analysis
Description