Machine learning ground motion models for a critical site with long-term earthquake records

Abstract

Ground motion models derived using data from widespread stations and earthquakes induce significant aleatory variability when applied to predict ground motions for a single site. This issue can be even worse when the sites of interest show complex site response characteristics and do not have enough similar sites in the ground motion model development dataset. This work provides a machine-learning-based ground motion modeling framework for sites with numerous earthquake recordings. The Onahama Port Array, a liquefiable site in Japan that has recorded more than 1200 earthquakes in the last 25 years, was selected as a case study. The gradient boosting method was employed to predict response spectra with periods from 0.01 to 10 s for earthquakes with Japan Meteorological Agency (JMA) magnitude ranging from 2.4 to 9.0. Different predictor combinations were tested, and two gradient boosting models were recommended: the basic model that requires earthquake magnitude and epicentral distance as inputs, and the optimal model that requires three additional inputs, focal depth, azimuth, and rainfall. The basic and optimal gradient boosting models have the root mean square error of 0.004 and 0.001 g, average r2 of 5-fold cross validation of 0.915 and 0.983, respectively. The results shed light on seismic hazard assessment for critical sites with long-term earthquake recordings.