Application of Focal Plane Directions for Estimating Ground Motion Models with General Regression Neural Networks

The general regression neural network (GRNN) allows testing of various events, wave paths, and site features as ground motion model (GMM) predictors. The GRNN is determined by all previous measurements, while this research aims to find the best conditional probability distribution f(X vertical bar y) of ground motion predictors given ground motion values. The f(X vertical bar y) is estimated by a kernel estimator. The GRNN was modified. Instead of the Euclidean distance of all predictors, we look for various metric spaces of predictors that minimize cross-validation error. This approach was used in the search for the GMM whose predictors include values describing focal mechanisms. Nodal plane angles were applied. GMMs containing different predictor configurations were compared. It was noticed that rake and strike nodal plane angles have an impact on the GRNN GMM while the dip angle does not. The databases of ground motions and events containing focal mechanisms, NGA-West2 and from Lubin Glogow Copper District in Poland, were used for the research.