Towards fully data driven ground-motion prediction models for Europe

We have used the Artificial Neural Network method (ANN) for the derivation of physically sound, easy-to-handle, predictive ground-motion models from a subset of the Reference database for Seismic ground-motion prediction in Europe (RESORCE). Only shallow earthquakes (depth smaller than 25 km) and recordings corresponding to stations with measured V s 30 properties have been selected. Five input parameters were selected: the moment magnitude M W , the Joyner–Boore distance R J B , the focal mechanism, the hypocentral depth, and the site proxy V S 30 . A feed-forward ANN type is used, with one 5-neuron hidden layer, and an output layer grouping all the considered ground motion parameters, i.e., peak ground acceleration ( PGA ), peak ground velocity ( PGV ) and 5 %-damped pseudo-spectral acceleration ( PSA ) at 62 periods from 0.01 to 4 s. A procedure similar to the random-effects approach was developed to provide between and within event standard deviations. The total standard deviation ( σ ) varies between 0.298 and 0.378 (log 10 unit) depending on the period, with between-event and within-event variabilities in the range 0.149–0.190 and 0.258–0.327, respectively. Those values prove comparable to those of conventional GMPEs. Despite the absence of any a priori assumption on the functional dependence, our results exhibit a number of physically sound features: magnitude scaling of the distance dependency, near-fault saturation distance increasing with magnitude, amplification on soft soils and even indications for nonlinear effects in softer soils.