4:00 PM - 4:15 PM
[SCG55-03] Development of sequential estimation method of earthquake fault plane based on ground-motion indexes using Bayesian optimization
Keywords:Earthquake fault plane, Ground-motion index, Bayesian optimization
It is important to understand the characteristics of a large earthquake immediately after its occurrence, both for academic and disaster mitigation purposes. Currently, the estimation of its hypocenter, magnitude, and fault mechanism is routinely and automatically done with on the assumption of a point source. Furthermore, a rectangular fault model based on the spatial distribution of geodetic data (REGARDS, Kamawamoto et al., 2017; GlarmS, Grapenthin et al., 2014; G-FAST, Crowell et al., 2016) and a line source model based on the spatial distribution of seismic amplitude (FinDer; B¨ose et al., 2018) have also been automatically conducted. In this study, we developed a new method to estimate a fault plane model of an earthquake based on the spatial distribution of ground-motion index.
As a link between the spatial distribution of ground-motion index and earthquake fault plane, we use the ground motion prediction equation (GMPE). GMPEs have been derived to predict ground-motion index at an arbitrary point from information such as earthquake magnitude and event coordinate. In order to consider the spatial finiteness of fault rupture in large earthquakes, instead of using the hypocentral distance based on the assumption of a point source, a distance based on the earthquake fault plane (e.g., shortest distance from source fault) is used in recent GMPEs. Here, we use the base model of GMPE for crustal earthquakes and the correction terms for the amplification of deep sedimentary layers and shallow soft soils in Morikawa and Fujiwara (2013). The target ground-motion index is JMA seismic intensity. The unknown parameters are latitude, longitude, and depth at the center of the fault plane, length and width of the fault plane, strike and dip degrees of the fault plane, and moment magnitude.
The fault plane information that reproduces the spatial distribution of observed ground-motion index is sequentially estimated. This inverse problem can be regarded as one of the black-box optimization problems because of its nonlinearity and the possibility of the discontinuity in solution distribution. The black-box optimization problem can generally be solved by grid search or random search. However, if there are many combinations of parameters, it takes a long time to reach the optimal value. To improve the effectiveness, we use Bayesian optimization. As a tool for Bayesian optimization, we used Optuna (Akiba et al. 2019), an open-source hyperparameter auto-optimization framework by Preferred Networks. Tree-structured Parzen Estimator (Bergstra et al. 2011, 2013) was used as the algorithm for Bayesian optimization.
Preliminary synthetic tests suggest that the newly-developed method can produce solutions close to the inputs. It was also found that the solution uncertainty is different among the unknown parameters, and that the estimation parameters related to the finiteness of the fault are not well determined when the earthquake magnitude is small.
As a link between the spatial distribution of ground-motion index and earthquake fault plane, we use the ground motion prediction equation (GMPE). GMPEs have been derived to predict ground-motion index at an arbitrary point from information such as earthquake magnitude and event coordinate. In order to consider the spatial finiteness of fault rupture in large earthquakes, instead of using the hypocentral distance based on the assumption of a point source, a distance based on the earthquake fault plane (e.g., shortest distance from source fault) is used in recent GMPEs. Here, we use the base model of GMPE for crustal earthquakes and the correction terms for the amplification of deep sedimentary layers and shallow soft soils in Morikawa and Fujiwara (2013). The target ground-motion index is JMA seismic intensity. The unknown parameters are latitude, longitude, and depth at the center of the fault plane, length and width of the fault plane, strike and dip degrees of the fault plane, and moment magnitude.
The fault plane information that reproduces the spatial distribution of observed ground-motion index is sequentially estimated. This inverse problem can be regarded as one of the black-box optimization problems because of its nonlinearity and the possibility of the discontinuity in solution distribution. The black-box optimization problem can generally be solved by grid search or random search. However, if there are many combinations of parameters, it takes a long time to reach the optimal value. To improve the effectiveness, we use Bayesian optimization. As a tool for Bayesian optimization, we used Optuna (Akiba et al. 2019), an open-source hyperparameter auto-optimization framework by Preferred Networks. Tree-structured Parzen Estimator (Bergstra et al. 2011, 2013) was used as the algorithm for Bayesian optimization.
Preliminary synthetic tests suggest that the newly-developed method can produce solutions close to the inputs. It was also found that the solution uncertainty is different among the unknown parameters, and that the estimation parameters related to the finiteness of the fault are not well determined when the earthquake magnitude is small.