Development of a site-specific ground motion model with a machine learning technique

Ground motion models (GMM) have been conventionally 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 create a GMM specific to the location based on its records. In recent years, the application of machine learning to GMMs has become popular (e.g., Derras et al. 2012; Kubo et al. 2020), and can provide better prediction performance than equation-based models. In this study, we construct the site-specific GMM using a machine learning technique, and compare its prediction performance with that of the general GMM. For the original dataset, we referred to a prototype strong-motion database in Japan (Morikawa et al. 2020, JpGU-AGU). 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 were 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 periods between 0.05 and 10 seconds. Using the entire training data, a general GMM was constructed 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 (AVS30). We also constructed a site-specific GMM with three explanatory variables (moment magnitude, epicentral distance, and event depth) using the training data at a single station. For a machine learning technique, we adopted a random forest algorithm in scikit-learn. The results indicate that the site-specific GMM at a station with sufficient training data (> 1000) has a good performance in the prediction of not only training data but also test data compared to the general GMM. At a station with insufficient data (~100), the prediction performance of the site-specific GMM in the test data is much worse than that of the general GMM.