Machine learning-based ground motion models for predicting PSAs of borehole motions in Japan

Abstract

Numerous ground-motion models (GMMs) that predict the intensities of surface ground motions have been previously developed based on regression analysis (RA). This study develops GMMs to estimate 5% damped pseudo-spectral accelerations (PSAs) for 30 periods (0.01-7.0 s) for within-rock ground motions, based on machine learning (ML) methods (i.e., two ensemble methods (random forest (RF) and gradient boosting (GB)) and an artificial neural network (ANN)). GMMs are developed separately for four earthquake types (main and aftershocks of active crustal region events and those of subduction zone interface events), considering the differences in the characteristics of each earthquake type. We utilize 20,041 ground motions recorded at 575 borehole stations in Japan during 602 earthquakes with moment magnitudes greater than 5.0 and rupture distances shorter than 300 km. The prediction performances of GMMs based on RF, GB, ANN, and RA are evaluated by the standard deviations of the total, between-event, and within-event residuals. The GMMs based on the three ML methods (RF, GB, and ANN) perform better than the RA-based models. The RF-based GMMs resulted in the most accurate prediction of the PSAs of within-rock ground motions with a small bias and variance, which can enhance the seismic designs and seismic hazard assessments for underground structures.