Statistical application of deep learning-based generalized ground motion models – from probabilistic seismic hazard analysis to structural fragility analysis

This study presents a comprehensive implementation and evaluation of deep learning-based generalized ground motion models (GGMMs) applied across key stages of performance-based earthquake engineering (PBEE).Traditional ground motion models (GMMs) typically rely on assumed empirical relationships between earthquake characteristics-such as magnitude and rupture distance-and intensity measures (IMs) like peak ground acceleration (), spectral acceleration ( ), etc.They also treat the prediction of IMs independently, disregarding the internal cross-dependencies among them.These limitations, combined with their generally low predictive accuracies, have prompted the development of various data-driven machine learning-based GMMs.In this study, data-driven GGMM proposed by the authors for crustal ground motions respectively, is employed.The GGMMs are based on long short-term memory (LSTM) recurrent neural networks that capture amplitude-based (e.g., , ), frequencybased (e.g., peak ground velocity), energy-based (e.g., Arias intensity, cumulative absolute velocity), and duration-based (e.g., significant duration) IMs as a sequence of cross-dependent parameters.Using the earthquake characteristics and site parameters, such as soil shear-wave velocity, as inputs, the GGMMs output a cross-dependent IM vector.Using the earthquake rupture forecast (ERF) data, the GGMMs are applied to conduct probabilistic seismic hazard analysis (PSHA) for Los Angeles, representing crustal seismicity-yielding hazard curves.The hazard curves guide the selection and scaling of 30 ground motions used in multi-strip analysis for three distinct steel moment frames (SMFs) across 10 hazard levels.Nonlinear time-history analyses are performed to derive fragility curves for each SMF, quantifying their vulnerability under different seismic conditions.This process is replicated with conventional GMMs and the conditional mean spectrum (CMS) approach.The results obtained at each stage, i.e., hazard curves, hazard spectra, selected motions, and SMF seismic demands from the multi-strip analysis, are statistically compared between the conventional GMM and deep learning-based GGMM approaches.The findings underscore the value of GGMMs in both hazard estimation and structural response assessment and finally provide a complete framework for the deployment of the machine learning-based GMMs for PBEE.