A recurrent-neural-network-based generalized ground-motion model for the Chilean subduction seismic environment

Abstract

This paper proposes a deep learning-based generalized ground motion model (GGMM) for interface and intraslab subduction earthquakes recorded in Chile. A total of ∼7000 ground-motion records from ∼1700 events are used to train the proposed GGMM. Unlike common ground-motion models (GMMs), which generally consider individual ground-motion intensity measures such as peak ground acceleration and spectral accelerations at given structural periods, the proposed GGMM is based on a data-driven framework that coherently uses recurrent neural networks (RNNs) and hierarchical mixed-effects regression to output a cross-dependent vector of 35 ground-motion intensity measures (denoted as IM). The IM vector includes geometric mean of Arias intensity, peak ground velocity, peak ground acceleration, and significant duration (denoted as Iageom, PGVgeom, PGAgeom, and D5-95geom, respectively), and RotD50 spectral accelerations at 31 periods between 0.05 and 5 s for a 5 % damped oscillator (denoted as Sa(T)). The inputs to the GGMM include six causal seismic source and site parameters, including fault slab mechanism, moment magnitude, closest rupture distance, Joyne-Boore distance, soil shear-wave velocity, and hypocentral depth. The statistical evaluation of the proposed GGMM shows high prediction power with R2 > 0.7 for most IMs while maintaining the cross-IM dependencies. Furthermore, the GGMM is carefully compared against two state-of-the-art Chilean GMMs, showing that the proposed GGMM leads to better goodness of fit for all periods of Sa(T) compared to the two considered GMMs (on average 0.2 higher R2). Finally, the GGMM is implemented to select hazard-consistent ground motions for nonlinear time history analysis of a sophisticated finite-element model of a 20-story steel special moment-resisting frame. Results of this analysis are statistically compared against those for hazard-consistent ground motions selected based on the conditional mean spectrum (CMS) approach. In general, it is observed that the drift demands computed using the two approaches cannot be considered statistically similar and the GGMM leads to higher demands.