Machine learning–based ground motion models for shallow crustal earthquakes in active tectonic regions

Abstract

Data-driven ground motion models (GMMs) for the average horizontal component from shallow crustal continental earthquakes in active tectonic regions are derived using a subset of the Next Generation Attenuation (NGA)-West2 data set, including 14,518 recordings out of 285 earthquakes recorded at 2347 different stations. We use four different nonparametric supervised machine learning (ML) algorithms including Artificial Neural Network (ANN), Kernel-Ridge Regressor (KRR), Random Forest Regressor (RFR), and Support Vector Regressor (SVR) to construct four individual models. Then, we use a weighted average ensemble approach to combine these four models into a robust model to predict various ground motion intensity measures such as peak ground displacement (PGD), peak ground velocity (PGV), peak ground acceleration (PGA), and 5%-damped pseudo-spectral acceleration (PSA). The model input parameters are moment magnitude, rupture distance, V_S30, and Z_TOR. The ensemble modeling attempts to remove the drawbacks or deficiencies of different ML algorithms while capturing their advantages and accounts for epistemic uncertainty. Although no functional form is provided, the model can capture salient features observed in ground motions such as saturation as well as geometrical spreading, anelastic attenuation, and nonlinear site amplification. The response spectra and the magnitude, distance, V_S30, and Z_TOR scaling trends are consistent and comparable with the NGA-West2 GMMs including several additional input parameters. We used a mixed-effects regression analysis to split the total aleatory uncertainty into between-event, within-station, and event-site–corrected components. The model is applicable to magnitudes from 3.0 to 8.0, rupture distances up to 300 km, and spectral periods of 0 to 10 s.