Towards site-specific ground motion estimates in Greece using a partially non-ergodic, mixed-effects, neural network approach

Abstract

This study presents a new, data-driven, region-specific ground motion model for Greece. This model utilizes a neural network approach that eliminates the need for any a priori functional form. Due to limitations in the recent Greek dataset, selected records from the RESORCE database have been incorporated. A fully connected multilayer perceptron is employed to predict several ground motion intensity measures (GMIMs), including peak ground velocity (PGV), peak ground acceleration (PGA), and the 5 % damped pseudo-spectral acceleration (PSA) at 18 periods ranging from 0.01 to 4.00 s, for active shallow crustal earthquakes (h $\le$ 30 km). Given the available dataset information, this GMM is driven by three input parameters; moment magnitude (M_w), Joyner-Boore distance, R_JB (km), and the average seismic shear-wave velocity of the uppermost 30 m at the station site, V_S30 (m/s). Additional source parameters, such as focal mechanism and depth, were also tested. The linear mixed-effects algorithm of the lme4 package [1] is used to decompose the total ground-motion aleatory variability (GMAV) into inter-event residuals (δBe) and Site-to-Site residuals (δS2S) while analyzing their dependence on the magnitude and distance (heteroscedasticity). The sensitivity of GMIMs predictions to various input parameters is also analyzed. Results indicate that combining the partially non-ergodic assumption (δS2S) with the heteroscedastic model significantly reduces GMAV, while these data-driven predictions exhibit physical trends consistent with classical GMMs. This new GMM enables site-specific predictions throughout Greece, provided sufficient on-site recordings exist to derive the site-specific term δS2S_s.