Comparative Evaluation of Machine-Learning Models and Recalibrated GMPEs for Ground-Motion Prediction

Accurate estimation of strong ground motion is important for seismic hazard assessment and for quickly evaluating earthquake impacts after an earthquake. In this study, data-driven ground-motion prediction models are developed using Japanese data to estimate peak ground acceleration (PGA), peak ground velocity (PGV), peak ground displacement (PGD), and spectral acceleration (SA) using machine-learning methods. Ensemble regression techniques, including Random Forest (RF), Extreme Gradient Boosting (XGBoost), and Light Gradient Boosting Machine (LightGBM), are trained using strong-motion records from the Kiban Kyoshin Network (KiK-net) and the Kyoshin Network (K-NET) collected between 1997 and
2025.
For comparison, PGA is also estimated using a conventional ground-motion prediction equation (GMPE). The functional form of Shoushtari et al. (2018) is adopted, and its coefficients are recalibrated using the same Japan dataset. The data are divided into training, validation, and testing sets, and model performance is evaluated using the coefficient of determination (R²), root mean square error (RMSE), mean absolute error (MAE), and logarithmic residuals. Additional analyses, such as observed-versus-predicted comparisons and residual trends with distance, magnitude, focal depth, and VS30, are carried out to assess model behavior and identify possible biases.
The Random Forest model shows performance comparable to the recalibrated GMPE, suggesting that both approaches effectively capture the key effects of magnitude, distance, and site conditions on ground motion in Japan. Although the overall accuracy is similar, machine-learning models provide added advantages, including data-adaptive learning, stable residual patterns, and flexibility in predicting multiple ground-motion parameters. Therefore, machine learning can be considered a useful complementary approach that improves the robustness and applicability of ground-motion prediction for seismic hazard assessment.