11:45 AM - 12:00 PM
[SCG51-10] Prototype site-specific ground motion model using machine learning
Keywords:Ground motion model, Site-specific model, Prediction of spectra information
Ground motion models (GMMs), which predict ground-motion intensity from information such as hypocentral distance and earthquake magnitude, have been conventionally built based on records at several observation stations so that they can be applicable to the prediction of an arbitrary point. On the other hand, if sufficient records are available at a single station, it is possible to build a GMM specific to the location based on its records. Recently, the application of machine learning to building GMMs has become popular (e.g., Derras et al. 2012; Kubo et al. 2020; Okazaki et al. 2021), which can provide better prediction performance than conventional equation-based models. Machine learning, which provides flexible and highly accurate prediction based on past data, is expected to be a good match for site-specific GMM. In this study, we developed the site-specific GMM using machine learning, and compared its prediction performance with the general GMM.
For dataset, we used a prototype of the unified strong-motion database in Japan (Morikawa et al. 2020). This database consists of ground-motion records of K-NET and KiK-net of NIED, site information based on K-NET, KiK-net, and J-SHIS of NIED, and information on earthquake source of JMA and F-net of NIED. From this flat file, we retrieved ground-motion records satisfying the following conditions: (1) 4.0 <= Mw <= 7.5, (2) hypocentral distance <= 300 km, (3) event depth <= 200 km, and (4) PGA >= 1 gal. The dataset was divided into training data recorded from 1997 to 2015, and test data recorded from 2016 to 2017. The target ground-motion intensity is 5% damped acceleration spectra that consist of 46 period points between 0.05 and 10 s. Using the entire training data, a general GMM was built with five explanatory variables: moment magnitude, epicentral distance, event depth, top depth to the layer whose S-wave velocity is 1,400 m/s at the site (Z1400), and average S-wave velocity up to a 30 m depth at the site (VS30). We also built a site-specific GMM with three explanatory variables (moment magnitude, epicentral distance, and event depth) using the training data at a single station. The random forest algorithm in scikit-learn was adopted as the machine learning method.
The results indicate that the site-specific GMM at a station with sufficient training data (> 1000) has a good performance in predicting training and test data compared to the general GMM. At a station with insufficient data, the prediction performance of the site-specific GMM in the test data is not good, and the residual variation of the site-specific GMM in test data is worse than that of the general GMM. In addition, Z1400 and VS30, which have often been used as site proxy in previous general GMMs, are not well representative of a site proxy for predicting spectral information, and their predictions have a period-dependent bias. On the other hand, no such period-dependent bias was found in the prediction of the site-specific GMM, suggesting that the site-specific GMM is able to learn the site amplification characteristics from past ground-motion data.
For dataset, we used a prototype of the unified strong-motion database in Japan (Morikawa et al. 2020). This database consists of ground-motion records of K-NET and KiK-net of NIED, site information based on K-NET, KiK-net, and J-SHIS of NIED, and information on earthquake source of JMA and F-net of NIED. From this flat file, we retrieved ground-motion records satisfying the following conditions: (1) 4.0 <= Mw <= 7.5, (2) hypocentral distance <= 300 km, (3) event depth <= 200 km, and (4) PGA >= 1 gal. The dataset was divided into training data recorded from 1997 to 2015, and test data recorded from 2016 to 2017. The target ground-motion intensity is 5% damped acceleration spectra that consist of 46 period points between 0.05 and 10 s. Using the entire training data, a general GMM was built with five explanatory variables: moment magnitude, epicentral distance, event depth, top depth to the layer whose S-wave velocity is 1,400 m/s at the site (Z1400), and average S-wave velocity up to a 30 m depth at the site (VS30). We also built a site-specific GMM with three explanatory variables (moment magnitude, epicentral distance, and event depth) using the training data at a single station. The random forest algorithm in scikit-learn was adopted as the machine learning method.
The results indicate that the site-specific GMM at a station with sufficient training data (> 1000) has a good performance in predicting training and test data compared to the general GMM. At a station with insufficient data, the prediction performance of the site-specific GMM in the test data is not good, and the residual variation of the site-specific GMM in test data is worse than that of the general GMM. In addition, Z1400 and VS30, which have often been used as site proxy in previous general GMMs, are not well representative of a site proxy for predicting spectral information, and their predictions have a period-dependent bias. On the other hand, no such period-dependent bias was found in the prediction of the site-specific GMM, suggesting that the site-specific GMM is able to learn the site amplification characteristics from past ground-motion data.