Machine-Learning-Based Surface Ground-Motion Prediction Models for South Korea with Low-to-Moderate Seismicity

Abstract

Ground-motion prediction models (GMPMs) have been developed to estimate seismic intensity considering earthquake magnitude, source-to-site distance, site condition, and so on. This study proposes GMPMs to predict 5% damped pseudospectral acceleration (PSA) for 27 periods ranging from 0.01 to 10 s in Korea, based on three machine-learning techniques (i.e., artificial neural network [ANN], random forest [RF], and gradient boosting [GB]). We use 1189 ground motions recorded at 50 surface stations during the 77 earthquakes with a local magnitude (ML) greater than 3.0, including the Gyeongju and Pohang earthquakes with ML of 5.8 and 5.4, respectively. We compare the performances of the three machine-learning-based models and the classical regression-based model in terms of the coefficient of determination (R2), root mean square error (RMSE), mean absolute error (MAE), standard deviation of residuals, and between-event and within-event residuals. The GB-based model shows the best performance. In addition, we analyze the working process of the GB-based model using variable importance and partial dependence (PD) plots. Among the five independent variables (ML, epicentral distance [Repi], average shear-wave velocity of the upper 30 m [VS30], focal depth, and slope angle) used in this study, ML and Repi are the most influential variables and show strong correlations with PSAs. We apply the GB-based model to three recent earthquakes larger than ML 3.0, and the model accurately predicts the PSAs at various stations. We also generate maps of estimated PSA (PSAest) values for the four periods (T = 0.01, 0.1, 1, and 3 s) for the scenario earthquake with an ML of 5.0. We provide a method for training the GB-based model using the Python library, which can enhance the ground-motion prediction not only in Korea but also worldwide, and an executable version of the validated GB-based model.

Recent advances in earthquake seismology using machine learning

“…An advantage of ML models over conventional GMPEs is that nonparametric ML methods can learn the functions of ground-motion models directly from data without assuming regression equations, resulting in a more accurate and useful predictor within the range of training data. Some studies have reported that ML prediction models have higher accuracy than the existing GMPEs (Khosravikia and Clayton 2021;Oana et al 2022;Seo et al 2022). Khosravikia and Clayton (2021) reported that when sufficient data are available, the ML prediction models provide more accurate estimates than the conventional linear regression-based method.…”

Machine-learning models to predict P- and S-wave velocity profiles for Japan as an example

“…We used Scikitlearn (Pedregosa et al, 2011) for the implementation of Gradient Boosting (GB) and Random Forest (RF) algorithms and the Tensorflow (Abadi et al, 2016) for the Artificial Neural Network (ANN) algorithm. Note that comparing these three algorithms is a popular practice in the field of machine learning-based studies (e.g., Krauss et al, 2017;Kim et al, 2020;Jun, 2021;Seo et al, 2022). These methods represent different types of machine learning algorithms and have been proven effective in handling complicated relationships within various datasets.…”

An Early Forecast of Long‐Period Ground Motions of Large Earthquakes Based on Deep Learning

“…In this study, we applied deep learning technologies for immediate forecasting of distant LP ground motions from seismic observations near the epicenter, anticipating the feasibility of a computing facility to analyze real‐time observation data. Previous studies based on deep learning have successfully predicted peak ground accelerations (PGA) and peak ground velocities (PGV) of distant target locations (e.g., Derras et al., 2012; Zhang et al., 2021) and the distribution of shaking intensity over a wide area (e.g., Kubo et al., 2020; Lilienkamp et al., 2022; Seo et al., 2022) using immediately estimated source parameters (magnitude (Mw), location, and depth) from observed waveforms. Research is underway to forecast the PGA, PGV (e.g., Jozinović et al., 2020, 2022), and velocity response spectrum (Taya & Furumura, 2022) of target locations from observed waveforms without estimating source parameters.…”