Generalized ground motion prediction model using hybrid recurrent neural network

Abstract

AbstractModern seismic hazard analysis utilizes various ground motion prediction equations (GMPEs) to estimate intensity measures (IMs) such as spectral acceleration (Sa), Arias intensity (Ia), significant duration (D5‐95), and cumulative absolute velocity (CAV) as scalar parameters. The use of these GMPEs leads to independent estimations of IMs for the same earthquake scenario. Since the IMs belonging to the same earthquake scenario are naturally correlated, the GMPEs used for characterization of seismic hazard are expected to incorporate such correlation structure. Particularly, Sa corresponding to different periods of a single‐degree‐of‐freedom oscillator is estimated by GMPEs that do not explicitly account for internal correlations between the Sa at various periods for the same ground motion. The current approach to incorporate such a correlation for Sa spectrum is through a post‐processing technique developed by Baker and Jayaram; their method relates the spectral accelerations of different periods using functional forms describing linear correlations. However, the proposed correlation functions can be further improved and extended to more accurately estimate the spectrum of near‐fault high‐magnitude ground motions using high order dependencies among the spectral accelerations. This study proposes a generalized ground motion prediction model (GGMPM) using a hybrid recurrent neural network (RNN) framework that can be used to estimate a 29 × 1 correlated vector (denoted as IM) of RotD50 Sa at 26 periods and geometric means of Ia, D5‐95, and CAV using a set of seismic source and site parameters as inputs. This is an improvement to the current body of knowledge because the high order dependencies between the individual components of IM are incorporated. A RNN framework is developed that estimates the median vector of IM. The discrepancy between the IM estimated using the RNN framework and the IM computed from recorded motions is further minimized using the covariance matrix adaptation evolution strategy (CMA‐ES), which is a non‐convex optimization method. The residuals of the RNN framework are used to construct the inter‐event and the intra‐event covariance matrices to account for the inter‐event and intra‐event variabilities of the ground motions. Hence, given the source and site parameters, the RNN‐framework returns a median prediction of the IM, which is then combined with estimated inter‐event and intra‐event covariance matrices to obtain the probabilistic estimation of IM. Furthermore, this GGMPM framework is compared against various currently used GMPEs, and the results of these comparisons demonstrate that the proposed GGMPM leads to improved predictions while maintaining the internal dependencies of the IM components.