Explainable XGBoost–SHAP Machine-Learning Model for Prediction of Ground Motion Duration in New Zealand

Abstract

Although ground motion duration significantly influences structural response, there is a lack of accurate prediction models for ground motion duration. Ground motion duration plays a vital role in structural response during an earthquake. Even a small magnitude event may cause severe structural damage if the duration of the earthquake is long. Thus, accurate estimation of ground motion duration is essential for structural seismic design and analysis. This article uses machine learning techniques to estimate ground motion duration for the New Zealand region. This paper investigates the use of emerging machine learning algorithms to address this critical issue using data from the New Zealand earthquakes database. The utility of the prediction models is also evaluated using numerous parameters related to filtering frequencies, fault dimensions, S-wave triggering flag, etc., apart from the traditional source, path, and site. Other parameters, e.g., the usable frequency range and the uncertainty of the available parameters, are also considered in evaluating the prediction of the considered machine learning models. Root mean squared error, along with the coefficient of determination (R2), is used to evaluate the performance of the machine learning models. The method with the least difference between actual and predicted values on the test set is presented for each duration metric available within the New Zealand database. Most importantly, the game theory-based SHapely Additive exPlanations (SHAP) are provided as to whether each independent variable would push the predictions toward higher or lower values. These explanations demonstrate the relative importance of the parameters within the strong motion database in the prediction of earthquake duration.