ABSTRACT
We apply three data-driven selection methods, log-likelihood (LLH), Euclidean distance-based ranking (EDR), and deviance information criterion (DIC), to objectively evaluate the predictive capability of 10 ground-motion models (GMMs) developed from Iranian and worldwide data sets against a new and independent Iranian strong-motion data set. The data set includes, for example, the 12 November 2017 Mw 7.3 Ezgaleh earthquake and the 25 November 2018 Mw 6.3 Sarpol-e Zahab earthquake and includes a total of 201 records from 29 recent events with moment magnitudes 4.5≤Mw≤7.3 with distances up to 275 km. The results of this study show that the prior sigma of the GMMs acts as the key measure used by the LLH and EDR methods in the ranking against the data set. In some cases, this leads to the resulting model bias being ignored. In contrast, the DIC method is free from such ambiguity as it uses the posterior sigma as the basis for the ranking. Thus, the DIC method offers a clear advantage of partially removing the ergodic assumption from the GMM selection process and allows a more objective representation of the expected ground motion at a specific site when the ground-motion recordings are homogeneously distributed in terms of magnitudes and distances. The ranking results thus show that the local models that were exclusively developed from Iranian strong motions perform better than GMMs from other regions for use in probabilistic seismic hazard analysis in Iran. Among the Next Generation Attenuation-West2 models, the GMMs by Boore et al. (2014) and Abrahamson et al. (2014) perform better. The GMMs proposed by Darzi et al. (2019) and Farajpour et al. (2019) fit the recorded data well at short periods (peak ground acceleration and pseudoacceleration spectra at T=0.2 s). However, at long periods, the models developed by Zafarani et al. (2018), Sedaghati and Pezeshk (2017), and Kale et al. (2015) are preferable.
A ground motion logic tree for seismic hazard analysis in the stable cratonic region of Europe: regionalisation, model selection and development of a scaled backbone approach
