Item

Abstract

While the growing number of seismic records enhances our understanding of ground motion, data from large earthquakes remain limited for fully supporting reliable ground-motion modeling. Efforts to integrate simulated and observed data show promise, but a quantitative framework for validation and guidelines for the use of simulated data is yet to be established. This paper addresses these challenges by developing and evaluating a hybrid data-based Ground Motion Model (GMM) using the latest generation of the CyberShake simulation dataset and the NGA-West2 observational dataset for Southern California. A GMM based on symbolic learning is proposed as the candidate equation to explore the properties of this hybrid data approach. After preprocessing the data, GMMs are constructed and compared across three scenarios: using only observed data, only simulated data, and a hybrid of both. The results show that the predicted median values from the GMM calibrated with simulated data align closely with those from the observed data. This study also demonstrates that residuals from all three types of GMMs conform to a lognormal distribution. However, the residual dispersion for simulated data is smaller than that for observed data. Moreover, the standard deviation of the hybrid model decreases progressively as the proportion of simulated data increases. This means that the simulations reproduce the average properties of the ground motion but underestimate the variability and the most severe site and source effects. Additionally, recommendations are provided for building future simulation databases and effectively combining simulated and observed data.