BROADBAND GROUND MOTIONS FROM PHYSICS-BASED SIMULATION AND EMBEDDING-INTERPOLATION MACHINE LEARNING

Broadband ground-motion predictions of scenario earthquakes are critical for seismic hazard assessment. Physics-based numerical simulations can take into account realistic source and site characteristics, but their accuracy can be limited to long-period components because of computational cost and the resolution of underground models. Therefore, machine learning approaches began to be applied to produce realistic broadband ground motions from simulated long-period waveforms with the aid of observed broadband waveforms . Okazaki et al. (2021) proposed a machine learning method and validated it on observational records. In this method, acceleration envelopes and the Fourier amplitude spectra (FAS) are transformed from long- to short-period waveforms using an embedding-interpolation method and an artificial neural network (ANN), respectively. The embedding-interpolation method consists of embedding pairs of long- and short-period envelopes into a low-dimensional latent space using an extension of t-stochastic neighbor embedding (t-SNE) and interpolating short-period envelopes based on neighbor relationships in the latent space. The Wasserstein distance is defined on the space of envelopes, which is essential for extracting time variation of acceleration amplitudes in observational records. The ANNs of FAS are trained on two datasets (records at all stations and those at individual stations) successively, which enables the ANNs to learn site-specific responses. In this study, we extend the method of Okazaki et al. (2021). Source and site parameters are added to input variables of the ANN of FAS. In particular, site labels are added as the one-hot vectors, which enables to represent site-specific responses using a single ANN without carrying out a two-step training in the previous method. We validate the extended method on simulation results of scenario earthquakes in the Kanto region, Japan.