Proposing an optimal input feature combination for estimating ground response spectra using machine learning: a study in New Zealand

Three machine learning (ML) models, which are Random Forest (RF), Extreme Gradient Boosting (XGB), and hybrid XGB-RF, were employed in this study using the New Zealand ground motion database, with 5% damped pseudo-spectral acceleration (PSA) as the target variable. The relative importance of 15 input features was assessed using four different approaches. Based on the resulting rankings, a sensitivity analysis was conducted to examine the direct impact of each input feature on the prediction accuracy of the models. An optimal combination (CB) consisting of the six most influential features was selected to develop the proposed PSA prediction models. All ML-based models demonstrated strong predictive performance, achieving the coefficient of determination (\({R}^{2}\)) values exceeding 85%, with the hybrid model consistently providing the highest \({R}^{2}\) across all period ranges. Mixed-effects regression was applied to evaluate model uncertainty, including an assessment of the single-site standard deviation. The proposed models were validated by comparing their predictions with those from established ground motion prediction equations (GMPEs) and observed recordings, showing a good agreement in PSA values. Furthermore, the application of these models to Japanese data, due to their similar tectonic environment, indicated their potential to serve as robust foreign prediction models in future regional ground motion studies.