Fully Data-driven non-ergodic ground-motion prediction models for low to moderate seismicity areas: using RESIF-RAP, ESM, RESORCE and NGA-West 2 data.

Abstract

Abstract The aim of this work is to develop a fully non-ergodic ground motion prediction model (FNE-GMPM) that provides functional forms (ffs) for each of the world's 13 regions. The ffs are derived from machine learning of a given dataset drawn from four databases: namely RESIF-RAP, ESM, RESORCE and NGA-West2. The machine learning is performed by the neural network approach whose explanatory parameters are the moment magnitude (MW), Joyner-Boore distance RJB, average shear wave velocity in the first 30 m VS30, nature of VS30: (measured or estimated) and the focal Depth. The model thus established estimates the ground motion intensity measures (GMIMs). These GMIMs are represented by the peak ground acceleration and the peak ground velocity PGA and PGV respectively, as well as 5 as well as the 13-period acceleration pseudo-spectra from 0.04 to 4.00 s (PSA) for a damping of 5%. The 13 regions subject of this study are distinguished by their epistemic uncertainties. The aleatory variability is considered as heteroscedastic depending on the MW and the RJB. The consideration of the non-ergodicity of the heteroscedasticity and using the machine learning approach leads to a significant reduction of the aleatory variability. This work makes it possible to have strong motions for regions with low and moderate seismicity, such as metropolitan France.