Geospatial Site Terms for Nonergodic Ground Motion Models

抄録

Site terms of conventional ergodic ground-motion models largely rely on unified site proxies, such as the time-averaged shear-wave velocity in the upper 30 m (V_S30) and basin depth, which have been shown to induce large uncertainties at specific sites. Recent ground-motion model developers have used the peak frequency (f_peak) derived from horizontal-to-vertical spectral ratios (HVSRs) as an individual site parameter in ergodic site-effects models for central and eastern North America (e.g., Hassani and Atkinson, 2016, 2018). In this work, we demonstrate that using multiple site terms derived from geospatial data is advantageous in nonergodic ground-motion modeling by better representing site-to-site variability. We integrate nonparametric machine learning techniques and geospatial variables to develop fully data-driven nonergodic ground-motion models (GMMs). A decision tree ensemble method (gradient boosting model, GBM) is employed to predict PGA, PGV, and 5%-damped PSA using the NGA-West2 database. We examine the predictive power of 24 globally-available geospatial proxies for ground-motion modeling.