9:00 AM - 9:15 AM
[SSS11-07] Development of data-driven strong motion model and common integrated database
Keywords:strong ground motion, GMPE, GMM, database, simulation, machine learning
Strong motion observations in Japan have a history of nearly 70 years. In particular, due to the strengthening of the strong motion observation system after the Great Hanshin-Awaji Earthquake Disaster in 1995, a huge amount of strong motion observation data have been collected by many institutions and private companies, including more than 800,000 records by the strong motion observation networks K-NET and KiK-net of NIED. Information on underground structure exploration has been accumulated, and the construction of a nationwide subsurface velocity structure model for strong motion calculation is in progress. Strong motion prediction models (GMPEs) based on regression analysis of observation records have been developed and have contributed to the progress of seismic hazard assessment. The problem with strong motion prediction by GMPE is that the ability to predict rare and inexperienced events such as very large earthquakes is not guaranteed by observation records. In order to overcome these problems, a strong motion prediction method by physics-based simulation (PBS) using an advanced subsurface velocity structure model has also been developed. It is becoming possible to supplement the lack of observation records with simulation data based on theoretical models. In this study, we try to construct strong motion big data including inexperienced events by fusing observation data and PBS data based on theoretical model. We aim to develop strong motion prediction models that have high predictive performance and meet the needs of the practical side by using data-driven analysis that utilizes a common integrated strong motion database.
Seismic hazard and risk assessment are indispensable for ensuring safety and security against earthquake disasters. In seismic hazard assessment, there is "uncertainty" due to the natural variability of earthquakes and the limits of our human perception due to infrequent events. It is important for society to properly consider the “uncertainty” of seismic hazards to function risk management in preparation for earthquake disasters. Probabilistic seismic hazard analysis (PSHA) is one of the methods for such seismic hazard analysis. From the PSHA point of view, GMPE is currently the most practical strong motion prediction model in the world. In the United States, a unified observation database for the development of GMPE has been constructed, and based on this, multiple GMPEs are studied under the same conditions by multiple developers. Differences among multiple GMPEs are treated as epistemic uncertainty in PSHA and weighted according to the performance of each GMPE. On the other hand, GMPEs in Japan are developed by each developer under their own data set and conditions. Therefore, inadequate performance comparison and insufficient evaluation of epistemic uncertainty among multiple GMPEs have become a practical problem. It has also been pointed out that regression analysis based only on specific datasets has low prediction performance for events not included in the dataset. These are partly due to the lack of a unified strong motion database for the development of strong motion prediction models in Japan. To solve these problems, we have held a strong motion DB working group for the purpose of constructing an observation DB for the development of highly practical GMPE, and have been proceeding with trial production of the observation DB and formulation of GMPE specifications. The working group consists of GMPE researchers in Japan, NGA participants in the United States, and corporate researchers involved in structural design and risk assessment.
GMPE is a model secured by observation facts, but rare events that should be predicted have not yet been observed. In order to deal with this simple and fundamental problem, this research aims to construct "strong motion big data" by fusing observation data and PBS data. The goal of this study is to propose multiple new strong motion prediction models (GMM) with high prediction performance by analyzing strong motion big data based on various methods including machine learning by multiple teams with different specialties, and extracting the information hidden behind the data beyond the knowledge of strong motion research so far. Specifically, formulation of GMM requirement specifications from the viewpoint of utilization, development of fusion method of observation data and PBS data, GMM development by various methods, quantitative evaluation of GMM variation, area prediction by spatial interpolation, evaluation of GMM performance and epistemic uncertainty. In this presentation, we will introduce the current status of our research.
Seismic hazard and risk assessment are indispensable for ensuring safety and security against earthquake disasters. In seismic hazard assessment, there is "uncertainty" due to the natural variability of earthquakes and the limits of our human perception due to infrequent events. It is important for society to properly consider the “uncertainty” of seismic hazards to function risk management in preparation for earthquake disasters. Probabilistic seismic hazard analysis (PSHA) is one of the methods for such seismic hazard analysis. From the PSHA point of view, GMPE is currently the most practical strong motion prediction model in the world. In the United States, a unified observation database for the development of GMPE has been constructed, and based on this, multiple GMPEs are studied under the same conditions by multiple developers. Differences among multiple GMPEs are treated as epistemic uncertainty in PSHA and weighted according to the performance of each GMPE. On the other hand, GMPEs in Japan are developed by each developer under their own data set and conditions. Therefore, inadequate performance comparison and insufficient evaluation of epistemic uncertainty among multiple GMPEs have become a practical problem. It has also been pointed out that regression analysis based only on specific datasets has low prediction performance for events not included in the dataset. These are partly due to the lack of a unified strong motion database for the development of strong motion prediction models in Japan. To solve these problems, we have held a strong motion DB working group for the purpose of constructing an observation DB for the development of highly practical GMPE, and have been proceeding with trial production of the observation DB and formulation of GMPE specifications. The working group consists of GMPE researchers in Japan, NGA participants in the United States, and corporate researchers involved in structural design and risk assessment.
GMPE is a model secured by observation facts, but rare events that should be predicted have not yet been observed. In order to deal with this simple and fundamental problem, this research aims to construct "strong motion big data" by fusing observation data and PBS data. The goal of this study is to propose multiple new strong motion prediction models (GMM) with high prediction performance by analyzing strong motion big data based on various methods including machine learning by multiple teams with different specialties, and extracting the information hidden behind the data beyond the knowledge of strong motion research so far. Specifically, formulation of GMM requirement specifications from the viewpoint of utilization, development of fusion method of observation data and PBS data, GMM development by various methods, quantitative evaluation of GMM variation, area prediction by spatial interpolation, evaluation of GMM performance and epistemic uncertainty. In this presentation, we will introduce the current status of our research.