Accounting for uncertainties of magnitude-and site-related parameters on neural network-computed ground-motion prediction equations

Ground-motion prediction equations (GMPEs) are used to express seismic intensity mea-sures as a function of source-, path-, and site-related parameters. Functional models are still widely used for their computation. Fully data-driven approaches have been recently proposed based on artificial neural networks (ANNs). However, the estimation errors of the predictor parameters (e.g., the magnitude and VS30) are generally not accounted for in the development of GMPEs. In the present study, the uncertainty in the magnitude-and site-related parameters is considered in the establishment of GMPEs by ANNs. For this, an algorithm is proposed based on the generalized least-squares principle applied to ANNs training. A simulated database is used to validate the approach and to demonstrate the effect of the input parameter uncertainties on the GMPEs. Finally, the proposed model is applied to the Reference database for seismic ground motion in Europe (RESORCE) data-base. Results show that the consideration of uncertainty in the magnitude-and site-related parameters can reduce the total GMPE uncertainties by 4%–16%, whereas the median predictions remain similar