Enhanced Seismic Ground Motion Modeling With Conditional Variational Autoencoder

Abstract

The current research focuses on creating a Conditional Variational Autoencoder designed for encoding and reconstructing 5% damped spectral acceleration ($S_a$). This model integrates parameters related to the characteristics of the seismic source, propagation path, and site conditions, utilizing them as conditional inputs through the bottleneck layer. Unlike conventional Ground Motion Models, which typically use these parameters in a deterministic fashion, our model captures complex, nonlinear interactions between these parameters and ground motion through a probabilistic framework. The model is trained on an extensive data set comprising 23,929 ground-motion records from both horizontal and vertical directions, sourced from 325 shallow-crustal events in the updated NGA-West2 database. The input parameters encompass moment magnitude ($M_w$), Joyner–Boore distance ($R_{\mathrm{JB}}$), fault mechanism ($F$), hypocentral depth ($H_d$), average shear-wave velocity up to 30 m depth ($V_{s30}$), and the direction of ground motion ($\mathrm{dir}$). To validate the model's reliability, both interevent and intraevent residual analyses are conducted, affirming its robustness and applicability. Furthermore, the model's performance is assessed through residual analyses. Thus, this study contributes to advancing techniques in ground motion modeling, specifically enhancing seismic hazard assessment and the reconstruction of ground-motion data.