Forecasting duration characteristics of near fault pulse-like ground motions using machine learning algorithms

Abstract

The duration characteristics of near-fault earthquake ground motions play a significant role in the dynamic response of a structure. Linear regression-based models are extensively used to forecast ground motions and duration parameters. However, such an approach fails to account for the complexity arising from the non-linear patterns in the data set. Nevertheless, implementing machine learning algorithms has the ability to uncover these unexplored patterns as well as the unique characteristics of ground motions comprised in the datasets. In this study, statistical relationships between several duration metrics and intensity measures of near-fault ground motions are evaluated using machine learning algorithms. Four different machine learning algorithms, namely Regression, Decision Tree, Support Vector machines, and Gaussian Process regression model are trained to determine the optimum model. All these machine learning models were examined using the selected database of 200 near-fault pulse-like ground motions, which was split into two parts, with 75% of data used for training and the remaining 25% for testing. The results indicate that the fine tree model for bracketed duration, stepwise linear regression model for uniform duration, and the exponential and rational gaussian process regression model for significant and effective duration, showed more accurate and reliable results as compared to other models.