Explainable data-driven prediction of response spectra for 1D site response analysis

Ground acceleration response spectrum prediction plays a crucial role in seismic design for engineering structures. Conventional one-dimensional site response analysis methods often fail to simulate complex sites and dynamic soil processes accurately and with strong uncertainty, resulting in notable discrepancies between computed and measured response spectra. Leveraging seismic records from 2428 ground motion measurements at 43 horizontal site stations in the KiK-net database, a predictive model BO-XGBR-SS was developed based on Extreme Gradient Boosting (XGBoost). By disregarding constitutive assumptions and utilizing bedrock input and site information as training features, ensemble learning algorithms (XGBoost and Random Forest) were employed for model training, complemented by Bayesian optimization based on Gaussian processes for hyperparameter tuning. Comparisons show XGBoost performs better, prompting further enhancement through a stratified sampling training strategy guided by site categories to mitigate potential feature imbalances. Mean square error (MSE), Coefficient of determination (R2), Pearson correlation coefficient (R), spectral acceleration residual variability and residual probability distribution were used as the evaluation parameters. The average prediction error of the proposed model is reduced by more than 30% compared to the equivalent linear and nonlinear methods. Furthermore, the matching accuracy of each response spectrum prediction is analyzed using Dynamic Time Warping (DTW) to combine normalized response spectra and typical records, and the proposed BO-XGBR-SS model performs stably under a variety of site conditions, overcoming the shortcomings of high-frequency underestimation and anomalous amplification of the long-period response spectra in the one-dimensional site-response analysis methods. The model's generalization ability was validated using recent seismic motion records as an external dataset. Feature interpretation using SHAP analysis aligned with existing knowledge, affirming the model's effective capture of correlations between features and seismic response.