Abstract
Most existing ground-motion models (GMMs) provide spectral acceleration (Sa) predictions only up to about 10 s, limiting their use for structures sensitive to motions at periods greater than 10 s. To address this, this study develops a record-fusion ground-motion modeling framework that covers 0.01–16 s, using 7804 collocated acceleration-velocity record pairs from 126 earthquakes in Southern California, USA. The framework integrates Bayesian optimization of fusion weights with staged GMM term search. The resulting GMM extends Sa prediction beyond 10 s and exhibits physically reasonable magnitude scaling, distance attenuation, and site-condition scaling over the modeled period range. Between-event, site-to-site, and remaining within-event residuals are generally centered near zero across all periods and show no clear systematic bias. Comparisons with four benchmark NGA-West2 GMMs indicate that, although the dataset is smaller than those used for the benchmarks, particularly for large-magnitude events, the developed GMM produces median predictions within 0.01–10 s, broadly consistent with the benchmark models. Further comparisons with observed Sa and representative earthquake cases show that predictions within 0.01–16 s remain consistent with observed trends and show no evident attenuation instability or divergence, confirming the effectiveness of the proposed framework and the stable predictive behavior of the resulting GMM across the modeled period range.

This publication has 30 references indexed in Scilit: