At regional scales, modeling the subsurface structure for physics-based numerical simulations poses a challenge, hence two of the most common modeling techniques are juxtaposed in this study. Simulations from both models are carried out using a seismic modeling code SW4 for two aftershocks of the 2007 Niigata Chuetsu-Oki earthquake and validated against recorded data from Kashiwazaki-Kariwa Nuclear Power Plant and three K-NET stations in the form of waveforms, response spectra, amplification factors and goodness-of-fit, which highlight the importance of geotechnical layers in capturing the site-specific response. Regional attenuation of ground motions from the two subsurface models is compared with recorded peak ground accelerations in the region and NGA-West2 ground motion prediction equations, revealing the realistic scattering pattern that can only be replicated by acknowledging a 3D subsurface structure. In conclusion, both the models have their individual merits and limitations that are reported after extensive qualitative and quantitative validations.

In this study, we proposed a method to simply predict the rise part of real-time seismic intensity time-series. First, the duration between the P-wave arrival and achieving 95% of the maximum value of real-time seismic intensity was taken to be the characteristic time of the rise part. Then, its prediction equation using hypocentral distance, moment magnitude, source depth, average S-wave velocity up to 30 m depth, and top depth to the layer whose S-wave velocity is 1,400 m/s as explanatory variables was developed by regression analysis based on strong-motion records of 41 earthquakes. In addition, an approximate function based on a logarithmic function was proposed to reproduce the time-series shape of the rise part. Combining the prediction equation and function, the rise part of real-time seismic intensity at any point could be predicted and reproduced. We verified the effectiveness and limitations of this prediction approach with records of the 2016 Kumamoto earthquake and the 2011 Tohoku earthquake.

We constructed ground motion evaluation models of peak accelerations and response spectra using supervised machine learning based on a strong motion database. The common logarithmic standard deviations of the ratios of the predicted values to the observed ones in our models were 0.18-0.21; the variation here is less than that in previous ground motion prediction equations. The generalizability of the models was tested on the data of three earthquakes that occurred after the earthquakes used for the training dataset. The results showed that the prediction accuracy decreased for earthquakes with features that were not included in the training dataset; however, the models with features based on prediction results using the previous ground motion prediction equation could compensate for the bias and lack of training data.