Combining ground motion prediction models for epistemic uncertainty minimization

Abstract

Each ground motion prediction equation (GMPE) provides different median ground motion measures and variances computed from a set of input parameters since the data set and methodology used to develop the GMPE vary. These differences are captured by the epistemic uncertainty that can be reduced by combining multiple models. We describe how to minimize the epistemic uncertainty by sensitivity testing on various combinations of four NGA-West2 GMPEs. The correlation levels among models are suggested based on the ranges of moment magnitude, site-to-source distance, site conditions, and selected sub-regions. The prediction errors are highly correlated at short periods among all models, whereas correlations are coarse at long periods. The optimized weight method which uses correlations between errors of models is the most effective to reduce the error variation comparing to other weighting methods. The use of optimized weight method using conditional weights, however, does not significantly further reduce the variation. © 2019 Associazione Geotecnica Italiana, Rome, Italy.