Abstract
Global ground-motion models (GMMs) may have unsatisfactory performance in predicting intensity measures (IMs) in specific regions, due to regional differences in seismological and geological conditions. This paper presents a hierarchical Bayesian approach for regionalizing global GMMs with partial non-ergodicity. Prior information, such as a GMM calibrated with substantial global data, is integrated with regional ground-motion data to derive posterior inferences for model coefficients and residual components. Compared to the mixed-effects regression method, hierarchical Bayesian updating can reasonably quantify uncertainty across varying regional data sizes, enabling statistically explainable forward predictions in data-scarce regions. To partially relax the ergodic assumption, the total residual is decomposed into between-event, site-to-site, and within-event single-site components, thus capturing repeatable site effects at individual strong-motion stations. The Bayesian approach is applied to predict spectral acceleration ( SA ) at periods from 0.01 to 10 s, peak ground acceleration ( PGA ), and peak ground velocity ( PGV ) of vertical ground-motion components in Turkey by regionalizing an extensively used global GMM. The results show that the updated GMM offers advantages in representing non-ergodicity, improving predictive accuracy, and enhancing applicability, compared to the global GMM and existing local GMMs. The global model tends to overestimate ground motion intensities especially in moderate earthquake magnitudes. Furthermore, the effects of prior variances and data quantity are investigated to provide insights into Bayesian modeling strategies. Also, the correlations among SA , PGA , and PGV for vertical, horizontal, and mixed component pairs in Turkey are quantified for the first time, thereby enabling vector-valued IM characterization for three-component ground motions.