Nonparametric ground motion models of arias intensity and significant duration for the Italian dataset

Abstract

Abstract In the field of ground motion simulation, the stochastic site-based methodology relies on the existing database of ground shaking. Based on these methodologies, several properties of seismic signals are used to simulate seismic waves. These parameters could be evaluated either parametrically via linear or nonlinear regression techniques or non-parametrically via sophisticated machine-learning algorithms. Nonetheless, parametric models, which consist of a particular mathematical formulation, can be a source of large bias. In this study, machine learning techniques are employed to develop predictive models for two main input parameters of a stochastic site-based ground motion model: Arias intensity and significant duration, which control the time variation of the simulated ground shakings. The Arias intensity, defined by the integral of the square of the acceleration time series, and the significant duration, which is related to the strong shaking phase of an earthquake, are also of particular interest in structural and geotechnical engineering fields. For this purpose, the random forest approach is employed to develop prediction models for the Italian database. To guarantee the prediction accuracy of the models also for unseen future data, only 80 percent of the data is used for training, and the rest is reserved for testing the trained model. The model hyperparameters are tuned to control bias and variance trade-offs by k-fold cross-validation. For each model, a set of hyperparameters is selected, and a possible range is given. Then, a Bayesian optimization technique is implemented to find the best set of these hyperparameters among the given range. All these models provided promising results compared to the prior models in the literature.