Quasi Real-Time Estimation of Response Spectral Distribution of Strong-Motions Using Machine Learning

Abstract

We investigate the surrogate model by machine learning for rapid estimation of response spectral distribution in earthquake disaster, which can take into account nonlinear site response and period-dependent spatial correlation of response spectra. We proposed a method to simulate the above behavior by combining interim outputs by spatial interpolation with linear soil amplification and deep learning networks. The results for the city of Sendai showed that the learned model is effective in estimating the response spectrum distribution considering the effect of nonlinear amplification for periods of about 1 second or less, and can be used as a surrogate model for the existing methods with an accuracy of about 0.1 in the normal log standard deviation. On the other hand, the accuracy of the intermediate output was higher for periods longer than about 1 second. The reason for the lower accuracy at longer periods may be the effect of surface waves or the period dependence of hyperparameters not considered in this study, and improving the accuracy at longer periods is an issue for future work.