Broadband Ground Motions from 3D Physics‐Based Numerical Simulations Using Artificial Neural Networks

Abstract

In this paper, a novel strategy to generate broad-band earthquake ground motions from the results of 3D physics-based numerical simulations (PBS) is presented. Physics-based simulated ground motions embody a rigorous seismic wave propagation model (i.e., including source-, path-and site-effects), which is however reliable only in the long period range (typically above 0.75-1 s), owing to the limitations posed both by computational constraints and by insufficient knowledge of the medium at short wavelengths. To cope with these limitations, the proposed approach makes use of Artificial Neural Networks (ANN), trained on a set of strong motion records, to predict the response spectral ordinates at short periods. The essence of the procedure is, first, to use the trained ANN to estimate the short period response spectral ordinates using as input the long period ones obtained by the PBS, and, then, to enrich the PBS time-histories at short periods by scaling iteratively their Fourier spectrum, with no phase change, until their response spectrum matches the ANN target spectrum. After several validation checks of the accuracy of the ANN predictions, the case study of the M6.0 Po Plain earthquake of May 29, 2012 is illustrated as a comprehensive example of application of the proposed procedure. The capability of the proposed approach to reproduce in a realistic way the engineering features of earthquake ground motion, including the peak values and their spatial correlation structure, is successfully proved.

Physics‐based probabilistic seismic hazard and loss assessment in large urban areas: A simplified application to Istanbul

“…45,43 Despite very popular, such a hybrid approach suffers of a twofold drawback which limits its use mainly for regional applications: first, the LF and HF parts of the resulting hybrid broadband waveform may lack correlation, because they are typically obtained by independent numerical approaches; second, while the spatial correlation of LF peak values of ground motion is ensured by the PBS, this may not occur for the HF part, since it is driven by stochastic approaches. For these reasons, we opted for the alternative approach already presented and validated in [Paolucci et al 54 ], referred to as ANN2BB, which takes advantage of ANN, generally used to estimate the nonlinear relationship between input and output variables for the correlation of which fast-and closed-form rules are not available. In our case, the input variables are the LF SA for ≥ * , where * is the threshold period corresponding to the range of validity of PBS, while the outputs are the HF SA.…”

Broad-band 3-D earthquake simulation at nuclear site by an all-embracing source-to-structure approach

“…Two aftershocks are simulated, by prepartitioning the computational domain over a distributed-memory supercomputer. Compared to Quinay et al [6], a extra mid-step is added herein: the seismic wave-field rendered by the regional scale analysis is enriched at HF (up to 30.0 Hz) by applying the so called ANN2BB hybrid procedure, introduced by [21] and based on the use of Artificial Neural Networks to predict the short-period (SP) part of the pseudo-acceleration response spectra Sa. This hybridization step is described in Section 2.3.3.…”

“…In this study, the teaching dataset is represented by the SIMBAD database [45], consisting of N db =501 three components high-quality accelerograms recorded world-wide, spanning a range of M W from 5 to 7.4 and epicentral distances less than 40 km. Two ANNs were iteratively trained upon this training set (refer to [21] for details on the training process and generalization features), one referring to the geometric mean of the horizontal components and one to the vertical one. In our case, the neural network is designed as a feed-forward two-layers Perceptron [46,47], featured by N h n =30 sigmoid hidden neurons and a linear output.…”

“…The number of nodes in the input layer N i n equals the number of input variables N LP Sa , whereas the number of nodes in the output layer N o n equals the number of target values N SP Sa . The ANNs were trained by exploiting the Levenberg-Marquardt algorithm [48,49] to perform back-propagation of the error and adjust the weights 4 (see [21] for further details).…”

“…To cope with these limitation, Paolucci et al[21] proposed to make use of Artificial Neural Networks (ANN), trained on a set of strong motion records, to predict the response spectral ordinates at short periods, using as input the LP ones obtained by the PBS (blue Sa spectrum inFigure 6a), and, then, to enrich the PBS timehistories by scaling iteratively their Fourier spectrum the ANN target spectrum (in small axes inFigure 6b). Further technical details on this so called ANN2BB procedure are outlined in[21].Compared to a standard hybrid approach, ANN2BB yields realistic waveforms, both in time and frequency domains, as well as it renders maps of short-period peak values of ground motion which reproduce more closely the coupling of source-related and site-related features of earthquake ground motion. The approach is suitable to portray in a realistic way the spatial correlation features of the peak values of ground motion although it is not suitable yet to obtain sets of waveforms with realistic spatial coherency features at high frequency.…”